WPS6842 Policy Research Working Paper 6842 Growth, Inequality, and Social Welfare Cross-Country Evidence David Dollar Tatjana Kleineberg Aart Kraay The World Bank Development Research Group Macroeconomics and Growth Team April 2014 Policy Research Working Paper 6842 Abstract Social welfare functions that assign weights to individuals in social welfare is due to changes in average incomes. based on their income levels can be used to document In contrast, the changes in inequality observed during the relative importance of growth and inequality changes this period are on average much smaller than changes in for changes in social welfare. In a large panel of industrial average incomes, are uncorrelated with changes in average and developing countries over the past 40 years, most incomes, and have contributed relatively little to changes of the cross-country and over-time variation in changes in social welfare. This paper is a product of the Macroeconomics and Growth Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at akraay@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Growth, Inequality, and Social Welfare: Cross-Country Evidence David Dollar (Brookings) Tatjana Kleineberg (Yale) Aart Kraay (World Bank) Keywords: growth, inequality JEL Classification Codes: O4, O11, I3 ddollar@brookings.edu, tatjana.kleineberg@yale.edu, akraay@worldbank.org. This paper is being prepared for Economic Policy. We would like to thank, without implication, David Bulman, Pete Lanjouw and Luis Serven for helpful comments. Financial support from the Knowledge for Change Program of the World Bank is gratefully acknowledged. The views expressed here are the authors' and do not reflect those of the Brookings Institution, the World Bank, its Executive Directors, or the countries they represent. 1. Introduction Concerns about inequality are at the forefront of many policy debates today. From the “we are the 99 percent” slogans of the Occupy Wall Street movement, to major policy speeches by US President Barack Obama, it is hard to escape the view that rising inequality poses major challenges in advanced economies. In the developing world too, much has been written about the adverse effects of high and rising inequality on the pace of absolute poverty reduction. Concerns about inequality appear to be pervasive beyond policy elites as well. Public opinion surveys suggest that strong majorities of respondents in advanced economies feel that the gap between rich and poor has been rising in recent years. According to a recent Pew survey, over 80 percent of respondents in advanced economies are of the view that the gap between rich and poor has been increasing. 1 These views no doubt in part reflect the fact that inequality has in fact been increasing in many countries. In the US over the past four decades, the Gini coefficient has risen from around 30 to around 40. Roughly the same has happened in China, only more rapidly: between 1990 and 2009 the Gini coefficient has increased from 32 to 42. Much of this increase has happened at the upper end of the income distribution. According to Atkinson, Piketty and Saez (2011), the income share of the top 10 percent in the US has increased from around 33 percent in 1970 to nearly 50 percent in 2007, while in China it increased from 17 to 28 percent between 1986 and 2003. However, it is also important not to lose sight of the fact that inequality has not increased in other countries, and has declined appreciably in still others. In the Atkinson, Piketty and Saez (2011) data, income shares of the top decile have been stable or even declining slightly since the mid-20th century in countries such as Germany, France, Switzerland, and the Netherlands. In Brazil, the Gini coefficient has declined from around 60 in the late 1990s to around 55 in the late 2000s. More systematically, in a large dataset of changes in inequality over periods at least five years long that we describe in more detail below, in roughly half of episodes the Gini coefficient of inequality increases, while in the other half of episodes it decreases. In this paper, we aim to shed light on a very simple question: how much do these changes in inequality, in either direction, matter? To answer this question we first need to be precise about what we mean by “matter”. Our approach here is unabashedly modest: we use standard tools of social welfare analysis to calculate how much more or less growth in average incomes a country would need 1 See Pew Research Center (2013) 2 over a given period in order to compensate for the observed change in inequality over the same period. We then document the size of this compensation, and its relationship to average income growth, in a large dataset of episodes of growth and changes in inequality covering 117 countries between 1970 and 2012. A simple example helps to illustrate our approach. The World Bank has recently made a major public commitment to the goal of promoting “shared prosperity”, defined as growth in average incomes of those in the bottom 40 percent of the income distribution in each country in the developing world. 2 As a matter of simple arithmetic, growth in average incomes in the bottom 40 percent is the sum of growth in average incomes, and growth in the share of income accruing to the bottom 40 percent. In China, for example, between 1990 and 2009 average incomes grew at 6.7 percent per year. At the same time, inequality increased in the sense that the income share of the bottom 40 percent declined from 20.2 percent to 14.4 percent, corresponding to an average annual rate of -1.7 percent per year. Combining these two observations, average incomes in the bottom 40 percent grew more slowly than overall average income, at 5 percent per year. From the standpoint of promoting shared prosperity, therefore, the growth “cost” of the increase in inequality in China over this period is about 1.7 percentage points of growth per year. Or put differently, had inequality not increased in China during this time, a growth rate of 5 percent per year (instead of the 6.7 percent that actually happened) would have generated the same improvement in social welfare according to the “shared prosperity” metric of the World Bank. This example illustrates the two key ingredients in our approach. First, in order to understand how much inequality changes “matter”, we need to specify a social welfare function that assigns weights to individuals in a country based on their incomes. In the case of the World Bank’s “shared prosperity” target, the implicit social welfare function weights everyone in the bottom 40 percent of the income distribution in proportion to their incomes, and assigns zero weight to everyone in the upper 60 percent of the income distribution. In our empirical work, we will consider a variety of very standard social welfare functions corresponding to different preferences over how income is distributed across individuals in a country. Second, growth in social welfare between two points in time will reflect growth in average incomes, and growth in the relevant inequality measure implied by the particular social welfare function under consideration. In the case of the World Bank’s “shared prosperity” target, the relevant inequality measure is the income share of the bottom 40 percent. For all of the social welfare 2 See World Bank (2013). 3 functions we consider, this decomposition will be additive, allowing for a straightforward quantification of the relative importance of growth and inequality changes, and the costs or benefits in terms of growth of the latter. In particular, the increase (decrease) in the relevant inequality measure in percentage points per year can be interpreted as the amount by which average income growth would have to be higher (lower) to deliver the observed growth rate of social welfare absent any changes in income distribution. We work with a large dataset of income distributions and average incomes covering 117 countries over the past four decades. This dataset combines the high-quality household survey data for developing countries underlying the World Bank’s global poverty estimates (Ravallion and Chen (2010)), with the Luxembourg Income Study (LIS) data for advanced economies. We focus on within-country changes in average incomes and income inequality observed over episodes at least five years long. In a sample of 285 such non-overlapping episodes, we calculate the contribution of growth in average incomes, and the contribution of changes in inequality, to growth in social welfare, for a wide variety of social welfare functions. Our basic findings are easily summarized. For all of the social welfare functions we consider, social welfare on average increases equiproportionately with average incomes. This reflects the fact that changes in the relevant inequality measures are not systematically correlated with changes in average incomes. For all but the most bottom-sensitive social welfare functions, the relationship between growth in social welfare and growth in average incomes is also quite precisely estimated. This reflects the fact that changes in inequality are small, in the sense that variation across episodes in inequality accounts for only a small fraction of the variation across episodes in changes in social welfare. Although changes in social welfare driven by changes in the relevant inequality measures are on average small and uncorrelated with growth in average incomes, it is nevertheless useful to understand their correlates. In particular, if there were some combination of policies and institutions that supported a given growth rate of overall per capita incomes, but in addition led to declines in the relevant inequality measure, then from the standpoint of promoting social welfare, such policies would dominate others that did not lead to the same declines in inequality. In the last part of the paper, we consider a range of variables identified as important for growth and inequality in the large empirical cross-country literature. In the spirit of comprehensive data description, we use Bayesian Model Averaging to systematically document the partial correlations between these variables and growth in average incomes and in inequality, and through these channels, their correlations with social welfare. 4 We find little in the way of compelling empirical evidence that any of these variables are robustly correlated with the relevant changes in inequality that matter for the set of social welfare functions that we consider. This paper builds on our previous work in Dollar and Kraay (2002) and Dollar, Kleineberg and Kraay (2013). In those papers we studied the relationship between growth in average incomes and growth in average incomes in the bottom 20 percent and bottom 40 percent of the income distribution. Our findings in this paper are generally consistent with this earlier work – in these papers we also found that changes in the income share of the poorest quintiles typically are small and uncorrelated with changes in average incomes. This paper expands on our earlier work by considering a much broader class of social welfare functions, which allows us to connect our previous findings specific to the income shares of the poorest quintiles with more general inequality measures. Our emphasis in this paper on decomposing changes in social welfare into a growth component and a distribution component is related to the large literature that has followed Datt and Ravallion (1992) in decomposing changes in poverty measures into growth and distribution components (see Kraay (2006) for a recent application to absolute poverty measures in a large cross-country dataset). In Section 2 we describe the social welfare functions we study, and show how growth in social welfare can be decomposed into growth and inequality changes. Section 3 briefly describes the data, and Section 4 contains our main results on the relative importance of growth and inequality changes for growth in social welfare. Section 5 uses Bayesian Model Averaging to systematically relate changes in social welfare to a variety of policy variables that have been considered in the cross-country empirical literature, and Section 6 concludes. 2. Growth, Changes in Inequality, and Social Welfare We study the relationship between growth in average incomes and growth in social welfare. Our starting point is a social welfare function which specifies the weights assigned to different groups in a country. While in principle we could work with social welfare functions defined over individuals, in practice data limitations imply that we only directly observe deciles of the income distribution in the developing countries in our dataset. For this reason we consider social welfare functions that assign 5 weights to deciles rather than individuals. In addition, we consider only social welfare functions defined over income. 3 In particular we consider social welfare functions which assign weights to deciles of the income distribution, indexed by = 1, … , , based on average income in these deciles: (1) W = (1 , … , ) where denotes average income in decile , with higher values of corresponding to richer deciles, i.e., 1 < 2 < ⋯ < . We will consider three well-known cases of this social welfare function. The first equally weights incomes in selected deciles, and assigns zero weight to incomes in the remaining deciles, i.e. (2) W=� ∑ =1 =1 where = 1 if decile matters for the social welfare function in question, and zero otherwise. For example, the World Bank’s shared prosperity goal mentioned in the introduction implies a social welfare function where 1 = 2 = 3 = 4 = 1 and zero otherwise, i.e. average incomes in the bottom 40 percent of the income distribution. As another example, the work of Piketty and Saez focusing on how incomes in the top decile have diverged from those in the remaining 90 percent of the income distribution can be interpreted through the lens of this social welfare function with 1 = 2 = ⋯ = 9 = 1 and 10 = 0, i.e average incomes in the bottom 90 percent of the income distribution. The second is a social welfare function proposed by Sen (1976), which weights the income of each group by its rank in the income distribution: 3 More generally, one could consider social welfare functions that reflect the lifetime utility of individuals, which in turn depends on present and future consumption, labour supply, and longevity, among many other things. See for example Jones and Klenow (2011), Basu et al (2013) for cross-country applications in a representative agent setting, and Perri and Krueger (2003) for an individual-level application in the US. 6 2 (3) W = 2 �( + 1 − ) =1 In contrast with the “selected deciles” approach, this social welfare function assigns positive weight to all deciles, and not just the poorest. However, in contrast with simple average income, this social welfare function is pro-poor in the sense that it places more weight on the incomes of poorer groups: the weight placed on incomes in the poorest decile ( = 1) is ten times the weight placed on incomes in the richest decile ( = 10). When the number of groups is large, Sen (1976) shows that = (1 − ), where represents average income and represents the Gini coefficient. This formulation will be particularly useful because it makes it very straightforward to interpret the magnitude of changes in (one minus) the Gini coefficient in terms of growth in average incomes. The third social welfare function is the class proposed by Atkinson (1970): 1 1− (4) 1 W = � � 1− � =1 1/ for ≥ 0 and ≠ 1 (for = 1 this limits to a geometric average of incomes, = �∏ =1 � ). To interpret this social welfare function, note that 1− is a weighted average of the incomes of each decile, with weights proportional to − . When = 0, the income weights are equal and the social welfare function collapses to average income. As becomes larger, the social welfare function assigns greater weight to the incomes of the poorest groups. This social welfare function can also be written as the product of a particular inequality measure and average incomes, i.e. = (1 − ), where 1 1 1 1− 1− denotes the Atkinson inequality index ≡ 1 − � ∑ =1 � . Figure 1 provides a graphical summary of these different social welfare functions. The top panel depicts the weights assigned to incomes of different percentiles of the income distribution, normalized to sum to one for each social welfare function. All of the social welfare functions are (weakly) downward sloping, reflecting the fact that they assign greater weight to lower incomes. As noted above, average incomes in the bottom 40 percent equally weights incomes in the bottom 40 percent but gives zero weight to incomes above this threshold. The Sen index weights incomes in proportion to ranks in the income distribution, and so is a straight downward-sloping line. Relative to the Sen index, the A(1) and A(2) measures assign even more weight to the poorest percentiles. Finally, we also show A(0) which is simply average incomes – this naturally weights incomes of all percentiles equally. 7 The bottom panel of Figure 1 shows the weights assigned to individuals in each percentile of the income distribution, in contrast with the weights assigned to their incomes in the top panel (again we normalize the weights to sum to one across percentiles). Consider first average incomes, as captured by A(0). Since average income weights all incomes equally, it assigns less weight to poor people and more weight to rich people, with weights proportional to relative incomes. The same is true within the bottom 40 percent of the income distribution for the social welfare function corresponding to “shared prosperity” – it assigns greatest weight to individuals at the 40th percentile of the income distribution, and lower weights to poorer people within the bottom 40 percent. This same non-monotonic set of weights on individuals is implied by the Sen index, which gives greatest weight to individuals near the middle of the income distribution, and low weights to those at the lower and upper extremes of the distribution. The A(1) measure is an interesting benchmark because it weights all individuals equally. For values of > 1, the Atkinson index assigns greater weight to poorer individuals. The main benefit of specifying an explicit social welfare function is that it allows us to quantify the contributions of growth in average incomes and changes in inequality to changes in overall social welfare. For the social welfare functions we consider, growth in social welfare is given by: (5) = � =1 where = is the elasticity of the social welfare function with respect to group ’s income. In order to distinguish between changes in average incomes and changes in relative incomes, it is useful to decompose income growth within each decile into average income growth plus the share of income accruing to that decile, i.e. = + , where denotes average income and denotes the income share of decile . Using this, combined with the fact that the social welfare functions we consider are all homogenous of degree one, i.e. ∑ =1 = 1, gives the following decomposition of growth in social welfare into growth in average incomes and a weighted average of changes in relative incomes. 4 4 The Sen index is homogenous of degree one only in the limit as N goes to infinity. By focusing on social welfare functions that are homogenous of degree one we are ruling out absolute inequality measures such as the Kolm 8 (6) = + � =1 Depending on the choice of SWF, this expression tells us how to weight changes in inequality happening at different parts of the income distribution, i.e. changes in the income shares of different individuals/groups within the population. In particular, (i) for the “selected deciles” social welfare (+1−) functions, the weights are = ∑ ; (ii) for the Sen index the weights are = ∑ ; and =1 =1(+1−) 1− (iii) for the Atkinson index the weights are = . These weights on changes in income shares �∑=1 1− � of different percentiles of the income distribution are exactly those shown in the bottom panel of Figure 1. This additive decomposition of welfare changes into growth in average incomes and a weighted average of changes in income shares allows a natural interpretation of changes in inequality in terms of growth. In particular, the second term in Equation (6) is just the difference between growth in welfare and growth in average incomes, and it can be interpreted as the “growth cost” of the observed increase in inequality, i.e. the amount by which growth would have to be higher in order to compensate for the increase in inequality, in the sense of delivering the same welfare growth. In the remainder of this paper we implement and analyze this decomposition in a large dataset of cross-country dataset of income distributions over the past forty years. 3. Data Our starting point is a large dataset of 963 irregularly-spaced country-year observations for which household surveys are available, covering a total of 151 countries between 1967 and 2011. This dataset is the merger of data available in two high-quality compilations of household survey data: the World Bank’s POVCALNET database, covering primarily developing countries, and the Luxembourg Income Study (LIS) database, covering primarily developed countries. The POVCALNET database is the dataset underlying the World Bank's widely known global poverty estimates. Its data on average index, as well as absolute poverty measures such as the headcount or the poverty gap. For such measures, decompositions such as the one suggested in Datt and Ravallion (1992) are more appropriate. 9 incomes and income distribution are based on primary household survey data. Roughly half of the surveys in the POVCALNET database report income and its distribution, while the other half report consumption expenditure and its distribution. When we study within-country changes in the distribution of income or consumption, we focus only on episodes where the initial and final survey are of the same type, i.e. both refer to income or both refer to expenditure. For terminological convenience, however, we will refer only to income. All survey means are expressed in constant 2005 US dollars adjusted for differences in purchasing power parity. For countries that are not covered in POVCALNET, we rely on the LIS database. 5 This expands our sample by adding 19 OECD economies. For these countries we construct mean income and decile shares directly from the micro data at the household level. The underlying surveys are nationally representative and intended to be comparable over time. We focus on the LIS measure of household total income, which is expressed in the raw data in current local currency units. We convert the survey means to constant 2005 USD and then apply the 2005 purchasing power parity for consumption from the Penn World Table, in order to be consistent with the POVCALNET data. Also for consistency with the POVCALNET data, we use LIS data on household size and equivalence scales to convert to average income and its distribution across individuals rather than households. Figure 2 gives an overview of the annual data availability from these two sources. LIS survey data starts earlier, going back to 1967, while POVCALNET observations start in the 1980s. Both databases have better country coverage in more recent years. For our empirical analysis, we organize the data into episodes or “spells”, defined as within- country changes in variables of interest between two survey years. Specifically, we calculate the average annual log differences of social welfare and its components, average annual growth in mean income and in the relevant inequality measures for each spell, recognizing that different spells cover periods of different length, depending on the availability of household survey data. We work with two sets of spells corresponding to different time horizons. The first is the longest available spell for each country, i.e. from the first available to the most recently-available survey for each country, while the second is the set of all possible consecutive non-overlapping spells lasting at least five years, beginning with the first available survey for each country. We can only compute these spells for countries where at least two comparable surveys separated by at least 5 years are available. 5 A handful of countries have surveys available both through POVCALNET and LIS. For these countries we use only the POVCALNET data, i.e. we do not switch within countries between POVCALNET and LIS. 10 While the data we rely on come from the most reliable cross-country datasets on income distribution currently available, there nevertheless are some spells where the changes in average income and/or changes in decile shares over the spell are extreme. To prevent these from unduly influencing our results, we clean the data by truncating the sample at the first and 99th percentiles of the distribution of average growth and of each of the 10 decile shares. This leaves us with a final sample consisting of 117 countries, with a median length between the first and last available surveys of 16 years. For the sample of spells at least five years long, we have a total of 285 spells, with a median spell length of 6 years. 4. Results 4.1 Basic Description We begin with some very simple descriptive analysis of the relative importance of growth and changes in inequality for growth in social welfare. The eight panels of Figure 3 plot growth in social welfare (on the vertical axis) against growth in average incomes (on the horizontal axis) in the sample of minimum 5-year long spells, for eight measures of social welfare: average incomes in the bottom 10, 20, 40 and 90 percent of the income distribution, the Sen index, and the Atkinson index with values of equal to 1, 2 and 3. Since growth in social welfare is the sum of growth in average incomes and the growth rate in the relevant inequality measure, the vertical distance between each point and the 45 degree line reflects the contribution of inequality changes to growth in social welfare. The striking feature of these graphs is that, for most social welfare functions, the data points cluster very closely around the 45-degree line. This indicates that changes in social welfare due to inequality changes (in either direction) are small, particularly in comparison with the large variation in growth rates in average incomes apparent on the horizontal axis of each graph. Only in the case of very bottom-sensitive social welfare functions that place a very high weight on the poorest deciles (such as average incomes in the bottom 10 percent, or the Atkinson A(3) measure), do we begin to see more dispersion around the 45- degree line. Figure 4 provides a complementary perspective on the relative size of changes in growth and changes in inequality. We first plot the distribution of growth rates across all episodes in our sample of spells at least five years long. This distribution has a mean of 1.5 percent per year, and very substantial variation: the standard deviation of growth in average incomes is 4.2 percent, and the 10th and 90th percentiles of the distribution of growth in average incomes are -2.8 percent and 6.2 percent 11 respectively. In short, growth in mean income is positive on average, and exhibits very large variation across spells. We then superimpose the distribution of growth rates of inequality changes in the same sample, for each of the eight social welfare functions we consider. The contrast between the distribution of growth rates and the distribution of inequality changes is stark. Consider for example the Sen Index, where the relevant inequality measure is (one minus) the Gini index. The mean growth rate of this inequality measure is very close to zero, at 0.0 percent per year. There is also substantially less variation across spells in changes in inequality than there is in growth: the standard deviation of changes in inequality for the Sen index is 1.2 percent, and the 10th and 90th percentiles of the distribution of inequality changes are -1.5 percent and 1.7 percent respectively. A useful thought experiment here is to ask the following question: from the perspective of having rapid expected growth in social welfare, would we prefer to take a random draw from the distribution of average income growth or a random draw from the distribution of inequality changes? If our social welfare function is the Sen index, the answer is unambiguously to prefer a draw from the distribution of growth rates. As noted above, the mean of the distribution of growth rates is 1.5 percent per year, while for inequality changes it is essentially zero. Even if we were to get a very good draw from the distribution of inequality changes (say at the 90th percentile), this would deliver a growth rate of social welfare only slightly faster than what we could get at the average of the distribution of mean income growth (1.7 versus 1.5 percent per year). These results on the relative importance of growth hold roughly the same for most of the social welfare functions we consider, with the exception of the most bottom-sensitive ones, average incomes in the bottom 10 percent, and the Atkinson A(3) index. For these social welfare functions, we continue to find that the distribution of the relevant inequality changes is centered on roughly zero: the average change in inequality across spells is 0.2 percent per year. However, changes in inequality exhibit considerably more variation. For example, for the Atkinson A(3) index the standard deviation of inequality changes is 4 percent, which is quite close to that for growth in average incomes. Table 1 reports summary statistics on the distributions of growth rates and inequality changes more systematically, and for different subsamples. The first column refers to average income growth, while the remaining eight columns correspond to growth in the relevant inequality measure for each of the eight social welfare functions we consider. Each horizontal panel corresponds to a different sample 12 of observations, and within each panel we report the number of observations, the mean, the standard deviation, and the 10th and 90th percentiles of the distribution of growth rates within each sample. The first panel reports summary statistics for the sample of long spells, and the sample of minimum-five- year spells. The second panel disaggregates the minimum-five-year spell sample into low-income, middle-income, and high-income countries. The third panel disaggregates the minimum-five-year spells by decades. 6 The summary statistics in the first panel reflect the previous discussion on the mean and relative variability of changes in average incomes and changes in inequality. Looking across income categories in the second panel, we see that growth on average was substantially higher in the low-income country spells (2.5 percent per year versus 1.3 percent per year in the middle- and high-income categories). Across all the different social welfare functions, however, we still see that growth in the relevant inequality measure was on average close to zero in all three income categories. Looking across decades, we see clear evidence of an acceleration in average growth rates of survey mean income, from near zero in the 1980s and 1990s to 2.8 percent per year in the 2000s. There was however little change in the growth rate of any of the inequality measures across decades, with the implication that average growth in social welfare increased by roughly as much as growth in average incomes increased. 4.2 Regression Analysis We next shed light on the joint distribution of growth rates and inequality changes, using a series of simple ordinary least squares regressions of growth in social welfare on growth in average incomes. Since the former is the sum of growth in average incomes and changes in the relevant inequality measure, the slope coefficient from this regression is: (7) 1 + COV � , � � / � � =1 6 A practical challenge for data description here is that only a small fraction of spells fall entirely within a single decade, and so it is not obvious how to assign the remaining spells to decades. To circumvent this problem, for each spell we define three variables measuring the fraction of years in the spell falling in each of three decades. For example, a spell lasting from 1989 to 1994 would have one-fifth of its years in the 1980s and four-fifths in the 1990s, and none in the 2000s. We then report weighted summary statistics by decade, weighting each spell by the fraction of observations falling in each decade. 13 i.e. it is 1 plus the slope coefficient of a regression of changes in inequality on growth. If the estimated slope coefficient is equal to one, this implies that on average, social welfare increases equiproportionately with average incomes. If, on the other hand, the slope coefficient is greater (less) than one, social welfare increases more (less) than equiproportionately with average incomes, reflecting a decline (increase) in the relevant inequality measure when average incomes increase. In addition, the goodness of fit of these regressions is informative about the relative importance of growth and inequality changes for growth in social welfare. In particular, it is straightforward to show that the share of the variance in growth in social welfare due to growth in average incomes is: (8) SD � � /SD � � where is the square root of the R-squared from a regression of growth in social welfare on growth in average incomes and () denotes the standard deviation of . 7 The top (bottom) panels of Table 2 report these regressions for the sample of spells at least five- years long (for the sample of long spells). As before, the eight columns correspond to the eight social welfare functions we consider. The regressions confirm the visual impressions from Figure 3 as the estimated slopes are all very close to one. In all cases we cannot reject (at the 5 percent significance level) the null hypothesis that the estimated slope coefficient is equal to one, indicating an absence of any statistically significant correlation between changes in average income and changes in inequality that are relevant for the social welfare measures we consider. One interesting exception is the slope for the income of the bottom 10%, which is 1.15 with a standard error of 0.08 in the sample of minimum- five-year spells, and is significantly greater than one at the 9 percent significance level. This constitutes some weak evidence that the income share of the bottom decile does on average increase when average incomes increase. However, we do not find this in the sample of long spells for this inequality measure, or for any of the other inequality measures. Turning to the variance decompositions, we see that, for most social welfare functions, most of the variation in growth in social welfare is due to growth in average incomes. For example, for the Sen Index and the income share of the bottom 40 percent, the shares of the variance of growth in social welfare due to growth in average incomes are 92 and 77 percent in the minimum-five-year spells. As we 7 We follow Klenow and Rodriguez-Clare (1997) in defining the share of the variance of + due to as ( () + (, )/ ( + ). Since the correlation between growth and inequality changes is small, the term (, ) plays a minimal role in the decompositions that follow. 14 move to measures that place more weight on the poorest deciles, this variance share declines further, to 60 percent for the bottom 20 percent measure, and 52 and 41 percent for the Atkinson A(3) and bottom 10 percent measures. In Table 3, we disaggregate our results by income level, focusing on the sample of spells lasting at least five years. The results are very similar for low-income, middle-income, and high-income countries. In all but one case we cannot reject the hypothesis that the slope is 1.0. The one exception is the social welfare function corresponding to income of the bottom 10 percent in middle-income countries, where the slope is slightly greater than one, at 1.16, indicating a weak tendency for the income share of the poorest 10 percent to rise as average incomes increase in this sample. We also continue to see that the share of the variance of growth in social welfare due to growth in average incomes is high, particularly for measures such as the Sen index and average incomes of the bottom 40 and 90 percent. This is particularly the case in high- and middle-income countries. In contrast, among low-income countries we find that the share of the variance of growth in social welfare due to growth is somewhat lower than in the middle- and high-income countries. In Table 4, we disaggregate our results by decade, again focusing on the sample of spells at least five years long. Across all periods and for all social welfare functions, we continue to find a slope coefficient that is very close to, and not statistically significantly different from one, indicating an absence of a significant relationship between changes in inequality and changes in average incomes. One consistent pattern, however, is that the share of the variance of growth in social welfare due to growth in average incomes declines slightly as we move from the 1980s to the 1990s to the 2000s. For example, for the bottom 40 percent social welfare function, this variance share declines from 85 percent in the 1980s to 72 percent in the 2000s. 8 In all of our results so far, we have relied exclusively on household survey means to construct growth rates in average incomes. A large literature has discussed substantial differences between growth in survey mean income and corresponding aggregates in the national accounts in some countries (see for example Deaton (2005) and Deaton and Kozel (2005) for the case of India in particular). 8 Interestingly, this decline in the variance share does not appear to be due to compositional effects (noticing that the sample size increases significantly over time). As a robustness check, we constructed a sample of 43 countries with one survey available near the middle of each of the three decades. We then constructed two sets of spells, from the mid-1980s to the mid-1990s, and from the mid-1990s to the mid-2000s. In this smaller set of spells we also find that the share of the variance of growth in social welfare due to growth in average incomes falls in the second set of spells relative to the first. 15 Without taking a stand on relative merits of national accounts versus household surveys as a measure of average living standards, we perform some simple robustness checks to see how our findings change if we rely on national accounts growth rates instead of household survey mean growth rates. The results are shown in Table 5, again focusing on the sample of spells at least five years long. Panel A repeats the results for all eight measures using the survey data. Panel B alternatively uses the growth rate of real private consumption from the national accounts. The slopes all continue to be very close to one. The main difference is that the share of welfare growth attributed to income growth declines modestly when we move to national accounts data. For example, for A(1) that share is 92% using the survey data; it declines to 85% using the alternative measure from the national accounts. 4.3 Why Is the Importance of Growth Different Across Social Welfare Functions? One striking feature of our empirical results is that the role of growth in accounting for changes in social welfare appears to be smaller for more bottom-sensitive social welfare functions. For example, in the variance decompositions reported in Table 2, we find that growth in average incomes accounts for just 41 percent of the variance of growth in incomes of the bottom 10 percent of the income distribution, but 77 percent for the social welfare function corresponding to growth in average incomes of the bottom 40 percent. Similarly, for the Atkinson Index with an inequality aversion parameter of = 1, growth accounts for 92 percent of the variation in growth in social welfare, but when the inequality aversion parameter increases to = 3, we find that growth accounts for just 52 percent of the variation. Mechanically, this finding reflects the fact that the growth rate of the income shares of the poorest deciles are the most volatile across spells in our dataset. This can be seen clearly in the top panel of Figure 5, which reports the standard deviation of the average annual growth rate of each decile share in the sample of spells at least five years long. The variance (across spells) of the growth rate of the income share of the bottom decile is 5.2 percent, while it just 1.4 percent for the fifth decile, and only 0.8 percent for the ninth decile share. To see why this matters for our variance decomposition, recall Equation (6) which decomposes growth in social welfare into growth in average incomes, and a weighted average of growth in the income shares of each decile. As we move to more bottom-sensitive social welfare functions, the weight on the lowest deciles increases. Since the growth rates of these lowest decile shares are the most volatile, this increases the variance of the second term in Equation (6), which in turn increases the share of the variance in growth in social welfare that is due to changes in 16 relative incomes. In short, social welfare functions that place greater weight on poorer individuals will be more responsive to variations in the income share of the poorest. At first glance, the finding of greater volatility in income shares of the poorest seems plausible, to the extent that the poor experience proportionately greater income shocks. However, in this section we show that greater volatility in the income shares of the poorest may also to some extent simply be a consequence of sampling variation, rather than reflecting any true differences in the population variance of income shares of the poorest. This in turn suggests that the share of the variance of growth in social welfare due to growth in average incomes might be understated in our results thus far in Table 2-Table 4, and particularly so for more bottom-sensitive social welfare functions. We illustrate this point by considering the thought experiment of drawing random samples of data on income at two points in time from a given population, which we can think of as the two endpoints of one of the “spells” we study in this paper. We assume the samples are drawn independently of each other (corresponding to household surveys that are repeated cross-sections of households, as is the case in most countries in our sample, rather than true panels that track individual households over time). We also assume that the population distributions of income from which we are sampling is lognormal (which is a reasonable approximation in most countries, see Lopez and Serven (2006)). Finally, we assume that the standard deviation of log income is the same for the two distributions. This provides us with a benchmark of what to expect when there is no change in inequality in the population over the spell. However, measured inequality, which can be summarized by the sample standard deviation of log incomes, will be different in the two samples, due to purely random sampling variation. In Appendix A we derive an analytical expression for the standard deviation of the growth rates of the income shares of different percentiles of the income distribution corresponding to this thought experiment, which we plot in the bottom panel of Figure 5. This pattern is remarkably similar to what we see in the actual data, with substantially higher volatility in the growth rate of the income shares of the poorest deciles, despite the fact that there are no changes in relative incomes between the two population distributions. This suggests that some of the apparent differences in the importance of growth in average incomes for growth in social welfare that we saw in Table 2-Table 4 may simply reflect the differential importance of sampling variation across decile shares. 17 We quantify this effect using the following illustrative calculation. The bottom panel of Figure 5 is based on the assumption that the household survey data comes from a simple random sample of = 5000 observations, and a standard deviation of log income of = 0.74, which corresponds to a population Gini coefficient of 40, which is the average observed value in our data. The variance of the growth rate of measured average income is the sum of the variance of true income growth, plus sampling variation, which under our assumptions is simply 2 2 /. 9 Based on this we can arrive at an 0.742 estimate of the true standard deviation of income growth, which is 4.0 percent, i.e. �0.0422 − 2 2500 = 0.0396, as opposed to the observed standard deviation which is 4.2 percent. Similarly, the observed standard deviation of the growth rates of the cumulative income shares of the bottom 10, 20 and 40 percent will reflect both true variation as well as well as sampling variation. Using the results in Appendix A, we can calculate the amount of sampling variation corresponding to our assumption of = 50000 and = 0.74, and subtract this from the observed variance in income shares to arrive at an estimate of their true standard deviations equal to 4.5 percent, 2.0 percent, and 1.6 percent for the income shares of the 10th, 20th and 40th percent (as opposed to the observed standard deviations of 5.2 percent, 3.4 percent, and 2.3 percent, respectively). Finally, we use these estimates of the true standard deviations of income growth and growth in income shares of the poorest to recalculate the share of the variance of growth in social welfare due to growth in average incomes. 10 This gives us larger variance shares, equal to 47 percent, 71 percent, and 86 percent for the income shares of the 10th, 20th and 40th percent (as opposed to the 41 percent, 61 percent, and 78 percent as reported in Table 2). To sum up, we have seen that the share of the variance of growth in social welfare due to growth in average incomes is smaller for more bottom-sensitive social welfare functions. This mechanically reflects the fact that growth in income shares of lower deciles exhibits greater variability in our dataset than those of higher deciles. However, we have seen that some of this greater variability in income shares of the poorest may simply reflect measurement error due to sampling variation. 9 The growth rate of average income is the difference between the mean of log income in the first sample and the second sample. The variance of mean log income is 2 /, and since the two samples are drawn indpendently, the variance of the difference in means is 2 2 /. 10 For this calculation, we assume that the true correlation between growth in the mean and changes in equality is the same as observed in the data. Since these observed correlations are small this assumption has little effect on our conclusions. 18 Adjusting for this suggests that the estimates of the share of the variance of growth in social welfare due to growth in incomes reported in Table 2-Table 4 are likely to be lower bounds. 4.4 More General Social Welfare Functions and Generalized Lorenz Dominance In all of our results thus far, we have relied on specific social welfare functions in order to be able to explicitly measure the contribution of inequality changes to growth in social welfare. In particular, the benefit of this approach is that it allows us to express inequality changes in terms of growth in average incomes, which in turn leads to straightforward variance decompositions. Although we have considered a variety of common social welfare functions, a drawback of this approach is that our conclusions do depend on the specific functional form of the selected social welfare function. And this in turn raises the possibility that there might be other social welfare functions for which the contribution of growth to improvements in social welfare is much smaller (or larger) than for the ones we have considered. To partially address this concern, we draw on the concept of generalized Lorenz dominance to determine the direction, although not the magnitude, of welfare changes for a much broader class of social welfare functions than those considered here, over the same set of spells. Shorrocks (1983) shows that as long as the social welfare function is increasing and concave in incomes (i.e. the social welfare function prefers more income to less, and less inequality to more), then social welfare is unambiguously higher if and only if the growth rate of cumulative average income of each percentile of the population is positive. In our case where we focus on decile grouped data, this corresponds to the case where growth in average incomes of the first decile, the first two deciles, etc. all the way up to growth in overall average incomes, are all positive over a given spell. If this is true, then social welfare will have improved over this spell for any social welfare function that satisfies the minimum requirements of being increasing and concave in incomes. To implement this, we divide our sample of spells into those with positive average growth rates and those with negative average growth rates. In the first group, we count the number of spells where the final distribution in each spell generalized-Lorenz-dominates the initial distribution. In the second group, we do the opposite, counting up the number of spells where the initial distribution (before the negative average growth experience) generalized-Lorenz-dominates the final distribution. In both cases, these correspond to cases where positive (negative) growth unambiguously raised (lowered) social welfare for any increasing and concave social welfare function. 19 The results of this calculation are shown in Table 6. Consider for example the 117 spells comprising the long spells sample. 96 of these spells correspond to periods of positive average growth, and of these, welfare unambiguously increases in fully 83 percent of spells. Conversely, for 57 percent of the 21 spells with negative average growth, welfare unambiguously declined. Overall, welfare changes in the same direction as average incomes in 78 percent of all spells. The small number of spells in which mean income and social welfare have not unambiguously moved in the same direction are generally ones in which the income growth rate is close to zero. 5. Policies, Growth, and Social Welfare So far we have seen that most of the variation in growth rates in social welfare across countries is due to cross-country differences in average growth. However, it may be important to understand the factors underlying the remaining variation in social welfare growth due to changes in inequality. In particular, if there were a combination of policies and institutions that resulted in the same growth rate of average incomes as some other combination of policies, but also delivered a reduction in inequality, then the inequality-reducing combination of policies would be preferable from the standpoint of improving social welfare. In this section we use cross-country regression analysis to document more systematically the relationship between a set of variables proxying for a variety of policy and institutional factors on the one hand, and growth in social welfare and its components on the other. Our starting point is the observation that each of the social welfare functions we consider is of the form = where is mean income and is a decreasing function of some inequality measure. For example, in the case of the Sen index, is one minus the Gini coefficient, which increases as inequality falls. Similarly, in the case of the income of the bottom 40 percent, is equal to the income share of the bottom 40 percent, which increases as inequality declines. For terminological convenience we will refer to as “equality”. As discussed in the previous section, growth in social welfare is the sum of growth in average income and growth in the relevant equality measure. We will consider a set of empirical models for growth in social welfare, of the following form: ′ ∆( , ) = ( , − 1) + ( , − 1) + (, ) + ( , ) (9) 20 In every model, growth in social welfare depends on lagged log-levels of average income and inequality, ( , − 1) and ( , − 1), to pick up any convergence in social welfare. Each model is defined by the specific set of right-hand-side variables ( , ) included in it, with indexing models. These variables are taken from a set of proxies for various aspects of the policy and institutional environment in the country. Naturally, the estimated coefficients on all right-hand-side variables, , , and , vary across models and therefore are also indexed by . This empirical specification allows us to decompose the contribution of various correlates of social welfare growth into their effects on growth in average incomes and growth in equality. In particular, Equation (9) above is the sum of a standard growth regression, as well as an analogous specification for changes in equality: ′ (10) ∆( , ) = ( , − 1) + ( , − 1) + ( , ) + ( , ) ′ (11) ∆( , ) = ( , − 1) + ( , − 1) + ( , ) + ( , ) where the estimated coefficients in the model for growth in social welfare are the sums of the coefficients from the models for growth in average incomes and growth in equality, i.e. = + ; = + ; and = + . We consider as explanatory variables a set of variables that have been identified as important correlates of growth in the large empirical cross-country growth literature. The growth correlates include a measure of financial development (M2 as percentage of GDP), the Sachs-Warner indicator of trade openness, the Chinn-Ito Index of financial openness, the inflation rate, the general government budget balance, life expectancy, population growth, the Freedom House measure of civil liberties and political rights, the frequency of revolutions, and a dummy variable indicating whether the country was party to a civil or international war in a given year. Most of these variables have been identified as important correlates of growth in one or more of three prominent meta-analyses of growth determinants (Fernandez, Ley and Steel (2001a), Sala-i-Martin (2004) and Ciccone et al. (2010)). We also consider some additional variables that have been found to be significant correlates of inequality in the much smaller existing cross-country literature on determinants of inequality. These consist of 21 primary school enrollment rates, a measure of educational inequality 11 (as emphasized by De Gregorio et. al. (2002)), and the share of agriculture in GDP (as emphasized for example in Datt and Ravallion (2002)). 12 Appendix B provides a detailed description of the definitions and sources of all of these variables. In each model, ( , ) includes one particular combination of these 13 variables, and we will consider all 213 = 8192 possible combinations of these variables. We rely on Bayesian Model Averaging (BMA) as a tool to summarize the results from this large number of regression models. In particular, BMA allows us to assign a posterior probability to each model, [ ], which indicates the relative likelihood of model compared with all of the other models considered. Intuitively, these posterior probabilities reflect a tradeoff between goodness of fit and parsimony: for two models with the same number of explanatory variables, the model that delivers the higher R-squared receives a higher posterior probability. Similarly, for two models with the same R-squared, the model with fewer explanatory variables receives a higher posterior probability. We calculate the posterior model probabilities for the regression for overall social welfare growth in Equation (9), and then use these to construct probability-weighted estimates of the slope coefficients across all models, i.e. 2 2 2 (12) E� � = � [ ] = � [ ] + � [ ] =1 =1 =1 This expression summarizes the estimated relationship between each of the explanatory variables and growth in social welfare, averaging across all models. Moreover, it allows us to separate the effects operating through growth and through increases in equality. Finally, for each variable we also calculate a posterior inclusion probability (PIP). This is simply the sum of the posterior probabilities of each model in which the variable appears, and is a useful summary of the relevance of that variable for growth in 11 Specifically, we use data on educational attainment by different levels of attainment from the Barro-Lee dataset to construct a (grouped) Lorenz curve summarizing the distribution of the total number of years of education across individuals, and from this calculate a corresponding Gini coefficient. 12 We also considered several other variables found to be significant correlates of inequality in some papers in the literature, but did not include them in our analysis because data coverage was very poor for many of the developing countries in our sample. These included indicators of labour market regulation and progressivity of tax systems (Checchi et. al. (2008)), public sector employment (Milanovic (2000) ), and social transfers (Milanovic (2000), De Gregorio et. al. (2002)). 22 social welfare. In particular, this measure identifies variables that appear in models that are more likely relative to other models. Table 7 summarizes our results. To conserve on space, we report results only for three social welfare functions: average incomes in the bottom 40 percent, the Atkinson Index with = 3, and the Sen Index. Our findings are qualitatively similar for the other social welfare functions considered in the paper. Our dataset consists of the sample of spells of changes in social welfare over periods at least five years long. Lagged log-levels of mean income and equality are measured at the beginning of each spell, while the remaining 13 explanatory variables are measured as averages over the spells. Our sample differs from the common practice in the cross-country empirical growth literature, which is to measure growth over regularly-spaced five- or ten-year intervals. The advantage of our approach of working with an irregularly-spaced panel is that it avoids the need to impute the household survey data across different years within a country, but instead relates actual changes in equality to explanatory variables observed over the same period. On the other hand, the disadvantage of this approach is that the coefficients on initial income and initial equality in Equations (10) and (11) are more difficult to interpret, because they reflect the rate of convergence over different time horizons. All of the regressions are estimated by ordinary least squares (OLS). 13 The first column in each panel reports the posterior inclusion probabilities for each variable. The remaining three columns report the estimated coefficients for growth in social welfare, growth in average incomes, and growth in equality, respectively. As discussed above, the slopes in the first column are the sum of the slopes in the remaining two columns, i.e. the overall effect of a given variable on growth in social welfare is the sum of its effects on growth in average incomes and on growth in equality. To aid in the interpretation of the slopes, we scale each one to show to the effect on growth (in percentage points per year) of a one standard deviation increase in the corresponding right-hand- side variable. Below each estimated coefficient, we report in parentheses the percentage of all models in which the estimated slope coefficient is statistically significant at the 95 percent level and is of the 13 Equations (10) and (11) are dynamic panel regressions, and the error terms are likely to include a country- specific time-invariance component (i.e. country effects). As such, they are subject to the usual concerns in the empirical growth literature about Nickell bias, as well as the usual concerns about potential endogeneity of other right-hand-side variables. Many papers in the empirical growth literature have sought to address these problems using the system-GMM estimator proposed by Arellano and Bond (1991). However, an under-appreciated concern with this approach is that the internal instruments based on appropriate lags and differences of explanatory variables are often weak in standard cross-country growth settings (see for example Bazzi and Clemens (2012)). We instead follow the recommendation of Hauk and Wacziarg (2010), who, following extensive Monte Carlo analysis of alternative estimators of growth regressions, conclude that pooled OLS performs best in practice. 23 same sign as the probability-weighted slope. Finally, note that the only difference across the three panels in the growth regression is the choice of the initial inequality measure, which is different for each social welfare function. As a result, the estimated coefficients on the remaining variables in the growth regression are in most cases quite similar across the three panels. Consider first the partial correlations between growth in social welfare and initial income and initial equality. We assume these variables appear in all specifications, so by construction the posterior inclusion probabilities are equal to one. The posterior probability-weighted estimated slopes in the social welfare regression are negative in all cases, and are highly significant in the vast majority of specifications, i.e. higher initial income levels, and higher initial equality, are significantly negatively correlated with subsequent growth in social welfare. To understand this finding, it is useful to consider separately the coefficients on these variables in the regressions for growth in average incomes, and growth in equality. Consistent with the large empirical growth literature, we find that lower initial income levels are associated with faster subsequent growth. The same is true of equality: lower initial levels of equality, i.e. higher initial inequality, are significantly associated with faster subsequent growth in equality, reflecting a tendency of equality to mean-revert. On the other hand, we find little evidence that higher initial levels of equality are significantly associated with subsequent growth, nor are initial income levels significantly associated with subsequent changes in inequality. Specifically, initial equality is significantly positively correlated with growth in only 4 percent of models (when the social welfare function is average incomes in the bottom 40 percent), and in 8 percent of models (when the social welfare function is the Sen index). This in turn casts some doubt on the frequently-heard concern that higher initial inequality might undermine subsequent growth. In fact, it is striking that even in the first and third panels, where the probability- weighted slope coefficient on initial equality is positive in the growth regression (indicating a positive partial correlation between greater equality and subsequent growth), the corresponding slope coefficient is strongly negative in the social welfare regression. This is because any beneficial effects of higher equality on subsequent growth in average incomes are offset by the tendency of equality to mean-revert, which reduces social welfare through an increase in inequality. Turning to the other variables in Table 7, some interesting patterns emerge. A few of them have consistently high posterior inclusion probabilities, indicating that they are important partial correlates of growth in social welfare in our sample. These include inflation (high inflation consistently is negatively correlated with growth in social welfare); population growth (faster population growth is associated 24 with slower growth in social welfare); political instability as measured by revolutions (greater instability leads to slower growth in social welfare); and the share of agriculture in GDP (a higher share is associated with slower growth in social welfare). Interestingly, for each of these variables, the main effect operates through the relationship with growth in average incomes: the coefficients in the third column of each panel (which measure the partial correlations with growth) are much larger in absolute value than the coefficients in the third column (which measure the partial correlations with changes in equality). Consider for example the relationship between inflation and growth in social welfare. A one standard deviation increase in inflation reduces average annual growth in social welfare by between 1.8 and 2.2 percent per year, depending on which social welfare function we consider. Nearly all of this effect comes through lower growth in average incomes, which declines by 1.7 percentage points. This estimated coefficient is significantly different from zero in all of the models in which it appears. 14 The probability-weighted slope coefficient from the equality regression is negative, suggesting that high inflation is disequalizing on average. However, the estimated effect is much smaller than the estimated effect on growth, and is rarely statistically significant. Another interesting case is the relationship between growth in social welfare and the share of agriculture in GDP, which is the one case among these four variables suggestive of a tradeoff. On the one hand, a higher share of agriculture in GDP is associated with substantially slower growth in average incomes, and this effect is significant in nearly all specifications. On the other hand, a higher share of agriculture in GDP is positively associated with subsequent growth in equality. However, this latter effect is once again much smaller than the effect on growth in average incomes, and moreover is never statistically significant at conventional levels. As a result, the growth effect dominates, and the effect of a larger share of agriculture in GDP on subsequent growth in social welfare is substantially negative. It is however noteworthy that the remaining variables mostly have quite low posterior inclusion probabilities, and rarely show up as significant correlates of either growth in average incomes or growth in the relevant inequality measures. This is at least somewhat surprising, given that these variables were selected for their prominence in the empirical growth literature. However, there are at least two important differences between the empirical specifications here, and those in the majority of papers in this broader literature. The first is that in our work, growth in average incomes is measured as growth in average income or consumption from the underlying household survey. As we have noted earlier, there are substantial differences between these growth rates and measures of mean growth taken from 14 This finding is in part driven by a fairly small number of spells with quite high inflation rates. 25 national accounts data. Second, the vast majority of papers in the empirical growth literature examine growth over regularly-spaced spells of fixed length, typically five or ten years long, and with a much larger set of observations. This contrasts with our sample, which is much smaller and irregularly-spaced, as dictated by the limited availability of household survey data in particular years. One way to assess the importance of these considerations is to re-estimate the results in Table 7, but replacing growth in household survey-based measures of average income with corresponding data from the national accounts. In particular, we use real per capita GDP growth from the national accounts, in order to be consistent with the empirical growth literature. However, we keep the structure of our irregularly-spaced panel of observations determined by the availability of survey data. For most variables, our findings are broadly similar to when we use household survey mean growth rates. In addition, for some of our social welfare functions, we find that the budget balance has a posterior inclusion probability greater than 0.5, and enters positively in the growth regressions, suggesting that higher budget deficits are correlated with slower growth. One further notable finding is that when we switch to national accounts growth rates, initial equality enters positively and significantly in roughly three-quarters of all specifications for the growth regression, suggesting that initial inequality is bad for subsequent growth. However, this does not imply that higher initial equality is good for growth in social welfare. This is because, as we have already seen in Table 7, higher initial equality is significantly correlated with subsequent declines in equality, and this effect is quantitatively larger than the positive effect on average income growth. As a result, the posterior probability-weighted slope coefficient on initial equality in the social welfare regression is still negative. Conclusions This paper is motivated by the widespread concerns about inequality and its consequences that are often heard in current policy discussions. Our objective in this paper is to provide some simple descriptive evidence on the implications of observed inequality for changes in social welfare. By specifying a social welfare function, it is possible to quantify the welfare effects of changes in inequality, and compare them with the welfare effects of growth in average incomes. In particular, we exploit the fact that growth in a number of standard welfare functions can be decomposed into growth in average incomes, and growth in a particular measure of inequality specific to each social welfare function. We implement these decompositions in a large dataset of high-quality data on income and its distribution across individuals, covering 117 countries over the past four decades. Using this data, we 26 construct a large set of non-overlapping episodes or spells of changes in average income and changes in its distribution that are at least five years long, and measure the contribution of growth and changes in inequality to growth in social welfare over each spell. Our basic findings are easily summarized. Most of the cross-country and over-time variation in changes in social welfare is attributable to growth in average incomes. In contrast, the contribution of changes in relative incomes to social welfare growth is on average much smaller than growth in average incomes, and moreover is on average uncorrelated with average income growth. These findings suggest that the welfare impacts of changes in inequality observed over the past four decades are small when compared with the welfare impacts of growth in average incomes. We have also seen that the relative importance of inequality changes for social welfare growth varies across the different types of social welfare functions we have considered. Specifically, the share of the variance of growth in social welfare due to growth in average incomes is smaller the greater is the weight that the social welfare function places on the poorest. For example, while the share of the variance of growth in social welfare due to growth in average income is 77 percent when the social welfare function is average incomes in the bottom 40 percent, it is 60 percent for the bottom 20 percent, and just 41 percent for the poorest 10 percent. Mechanically, this feature of our results is due to the fact that the observed income shares of the poorest exhibit considerably more volatility than the income share of those nearer to the middle of the income distribution. While some of this greater volatility in income shares of the poorest surely is real, we also demonstrate that at least some of it may simply be due to a differential effect of sampling variation on income shares at different points in the income distribution. An illustrative correction for this suggests that the share of the variation in social welfare growth due to growth in average incomes could be considerably higher even for social welfare functions that attach positive weight only to the very poorest. While the variation in the growth rate of most of the social welfare functions we consider that is due to changes in inequality is on average substantially smaller than the share due to growth, it is nevertheless potentially important to understand the factors underlying this variation. If, for example, there were some combinations of policies that generated the same rate of average growth as another combination, but at the same time reduced inequality, then the former combination might be preferable from the standpoint of social welfare growth. To investigate this we consider a set of variables intended to proxy for factors commonly thought to be conducive to growth in average incomes. We use Bayesian Model Averaging as a tool to systematically document the partial 27 correlations between these variables and social welfare growth. We find that the relationship between these variables and social welfare growth, to the extent that they are quantitatively important, comes mostly through their effects on growth in average incomes, rather than through changes in inequality. This of course does not imply that there are no policies that can influence inequality in ways that raise social welfare. However, it does suggest that the historical experience of a large set of developed and developing countries does not provide much guidance regarding the set of macroeconomic policies and institutions that might be particularly conducive to promoting growth in social welfare beyond their effects on aggregate growth. References Arellano, Manuel and Stephen Bond (1991). “Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Questions”. Review of Economic Studies. 58(2)277-297. Atkinson, Anthony, Thomas Piketty, and Emmanual Saez (2011). “Top Incomes in the Long Run of History”. Journal of Economic Literature. 49(1):3-71. Bazzi, Samuel and Michael Clemens (2013). “Blunt Instruments: Avoiding Common Pitfalls in Identifying the Causes of Economic Growth”. American Economic Journal: Macroeconomics. 5(2):152-186. Basu, Susanto, Luigi Pascali, Fabio Schiantarelli, and Luis Serven (2013). “Productivity and the Welfare of Nations”. Manuscript, World Bank and Boston College. Checchi, Daniele; Garcia-Penalosa, Cecilia. (2008). “Labour Market Institutions and Income Inequality”, Economic Policy issue 56, pp. 601-34, 640-49. Chen, Shaohua, and Martin Ravallion. (2010). “The Developing World Is Poorer Than We Thought, But No Less Successful in the Fight against Poverty,” The Quarterly Journal of Economics. Ciccone, Antonia and Marek Jacocinski. (2010). “Determinants of Economics Growth: Will Data Tell?,” American Economic Journal: Macroeconomics 2, 2:4, 222-246. Datt, Gaurav, Ravallion, Martin (1992). “Growth and redistribution components of changes in poverty measures: a decomposition with applications to Brazil and India in the 1980s”. Journal of Development Economics 38(2):275–295. Deaton, Angus, and Valerie Kozel. (2005). “Data and Dogma: The Great Indian Poverty Debate”, Oxford University Press 20:177-199. De Gregorio, Jose; Lee, Jong-Wha. (2002). “Education and Income Inequality: New Evidence from Cross- Country Data”, Review of Income and Wealth, v. 48, iss. 3, pp. 395-416. Dollar, David, and Aart Kraay. (2002). “Growth is Good for the Poor,” Journal of Economic Growth, 7, 195-225. 28 Dollar, David, Tatjana Kleineberg, and Aart Kraay (2013). “Growth Still Is Good for the Poor”. World Bank Policy Research Department Working Paper No. 6568. Feldkircher, M., & Zeugner, S. (2009). Benchmark Priors Revisited: On Adaptive Shrinkage and the Supermodel Effect in Bayesian Model Averaging. IMF Working Paper No. 09/202, International Monetary Fund. Fernandez, C., Ley, E., & Steel, M.F.J. (2001a). Model Uncertainty in Cross-Country Growth Regressions. Journal of Applied Econometrics, 16(5), 563-576. Fernandez, C., Ley, E., & Steel, M.F.J. (2001b). Benchmark prior for Bayesian model averaging. Journal of Econometrics, 100(2), 381-427. Hauk, William and Romain Wacziarg (2009). “A Monte Carlo Study of Growth Regressions”. Journal of Economic Growth. 14:103-147. Jaumotte, Florence, Subir Lall, and Chris Papageorgiou. (2013). “Rising Income Inequality: Technology, or Trade and Financial Globalization,” IMF Economic Review, v. 61, no. 2, pp. 271-309. Jones, Chad and Peter Klenow (2011). "Beyond GDP: Welfare Across Countries and Time". Working Paper, Stanford University. Klenow, Peter and Andres Rodriguez-Clare (1997). “The Neoclassical Revival in Macroeconomics – Has It Gone Too Far?”, in Ben Bernanke and Julio Rotemberg, eds. NBER Macroeconomics Annual. Cambridge, MIT Press, pp. 72-103. Kraay, Aart (2006). “When Is Growth Pro-Poor? Evidence from a Panel of Countries”. Journal of Development Economics. 80:198-227. Lopez, J. Humberto and Luis Serven (2006). “A Normal Relationship? Growth, Inequality, and Poverty”. World Bank Policy Research Working Paper No. 3814. Milanovic, Branko. (2000). “Determinants of Cross-Country Income Inequality: An 'Augmented' Kuznets Hypothesis”, Essays in honour of Branko Horvat, pp. 48-79. Perri, Fabrizio and Dirk Krueger (2003). "On the Welfare Consequences of the Increase in Inequality in the US". NBER Macroeconomics Annual. Pew Research Center (2013). “Economies of Emerging Markets Better Rated During Difficult Times”. http://www.pewglobal.org/2013/05/23/economies-of-emerging-markets-better-rated-during-difficult- times/. Sala-i-Martin, Xavier, Gernot Doppelhofer, and Roland I. Miller. (2004). “Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach.” American Economic Review, 94(4): 813-35. Sen, Amartya (1976). “Real National Income”. Review of Economic Studies. 29 Shorrocks, Anthony (1983). “Ranking Income Distributions”. Economica. February, pp. 3-17. World Bank (2013). “The World Bank Goals: End Extreme Poverty and Promote Shared Prosperity”. http://www.worldbank.org/content/dam/Worldbank/document/WB-goals2013.pdf. 30 Appendix A: Role of Sampling Variation In this appendix we detail the calculations underlying the bottom panel of Figure 5. We consider the case of sampling incomes from a lognormal distribution, with a fixed mean and standard deviation of log incomes given by and , respectively. The Lorenz curve for the lognormal distribution is given by (, ) = Φ(Φ−1 () − ), where ∈ [0,1] indexes percentiles of the income distribution and Φ(. ) is the standard normal cumulative distribution function. The income share of the ℎ percentile of the income distribution is: (, ) Φ−1 () (Φ−1 () − ) (A1) (, ) = = (Φ−1 () − ) = �Φ−1 ()� where (. ) denotes the standard normal probability density function and the last equality follows from applying the inverse function theorem. Given a random sample drawn at time from this lognormal distribution, and an estimate of the standard deviation of log income based on this sample, � , a natural estimate of the income share of the ℎ percentile of the income distribution at time is (, � ). Similarly, a natural estimate of the growth ℎ rate of the income share of the percentile between and − 1 is ∆ln�(, � )�, where ∆ is the difference operator. The variance of the growth rate of the income share of the ℎ percentile is given by � )�� = �ln�(, �∆ln�(, �−1 )�� = 2�ln�(, � )� − ln�(, �)�� (A2) 2 ln((, )) ≈ 2� � �] � [ � = The first equality above is just the difference operator. The second equality requires the additional assumption that the two random samples at and − 1 are drawn independently, i.e. the case of repeated cross-sections rather than true panels. The last expression is just the usual linearization, around some fixed point = �. To evaluate this variance, note that ln((, )) (Φ−1 () − ) (A3) = = Φ−1 () − where the first equality follows from the fact that only the numerator of the percentile share depends on , and the second equality exploits the functional form of the normal density function, i.e. ( ) ∝ − 2 /2. Inserting this into the previous gives the following expression for the variance of the growth rate of the income share of the ℎ percentile: (A4) �)2 [ � )�� ≈ 2(Φ−1 () − �∆ln�(, �] For a given amount of sampling variation in our estimate of the standard deviation of log income, Equation (A4) tells us how this is reflected in variability in the growth rate of the income share 31 of the ℎ percentile. The bottom panel of Figure 5 plots the theoretical standard deviation of the growth rate of the income share of the ℎ percentile, linearizing around the point � = 0.74, corresponding to a Gini coefficient 0.4, which is roughly the mean of all the Gini coefficients in our �] = /√2. We sample. The level of this series is arbitrary and depends on what we assume for � [ assume that = 0.74 , and to roughly match the data in the top panel of Figure 5 we assume a sample size of = 5000. While many of the household surveys in our dataset have much larger samples, it is worth keeping in mind that they rarely are simple random samples as is assumed here, and so the effective sample size is much smaller. To make the adjustments to the growth rates of the cumulative decile shares implied by this calculation, note first that, following the same steps as above, the variance of the growth rate of the ℎ cumulative decile share is: 2 (Φ−1 () − ) (A5) � )�� = 2�ln�(, �∆ln�(, �)�� ≈ 2 � �] � [ Φ(Φ−1 () − ) 32 Appendix B: Description of Explanatory Variables Used in Bayesian Model Averaging Variable Source Description / Adjustments Population WDI Population growth in percentage points growth Life expectancy WDI Life expectancy in years Financial depth; WDI Money and quasi-money (M2) as percent of GDP M2 as % of GDP Inflation rate WDI Inflation measure is calculated by taking annual log-differences from the WDI reported GDP deflator (local currency units). Budget balance WEO and data Data series on Budget Balance from Easterly, Levine, Roodman (2004) was used from Easterly, when available, after last available year, used WEO data. Levine, Roodman (2004) Revolution Cross-National Banks, Arthur S., Wilson, Kenneth A. 2013. Cross-National Time-Series Data Archive. Time Series Databanks International. Jerusalem, Israel; see http://www.databanksinternational.com Trade Openness Wacziarg- Wacziarg-Welch (2008) extension of the initial Sachs-Warner (1995) openness Welch (2008); measure is available through 2001. We update the series to 2010 using underlying extended data on tariffs, black market premium and export marketing boards. A country is through 2010. considered as closed if it has one of the following: Average tariff rates over 40 percent, black market exchange rate over 20 percent lower than the official http://www.a exchange rate, or a state monopoly on major exports (export marketing board). nderson.ucla.e 1. Tariffs: (Francis K.T. Ng “Trends in average applied tariff rates in developing and du/faculty_pa industrial countries, 1980-2006”; http://go.worldbnka.org/LGOXFTV550). No ges/romain.wa countries had tariffs beyond the 40 percent threshold at any time after 2000. cziarg/papersu 2. Black market premium: (Economic Freedom in the World 2012 report and m.html database from the Fraser Institute (http://www.freetheworld.com)). Data reports a 0-10 ranking where 10 implies no black market premium and 0 implies a premium of 50 percent or more. The black market premium is defined as the percentage difference between the official and the black market exchange rate. We assume that a score of 0-6 implies a premium of 20 percent or greater. 3. Export marketing board: In 2001 Wacziarg-Welch identified 12 countries as having an export marketing board based on various underlying data and sources. Clemens et al. update the classification through 2005, identifying three further countries has having liberalized or abolished their export marketing boards (Senegal (2002), Chad and Papua New Guinea (2005)). In our update we assume that none of the remaining 9 countries (Central African Rep, Congo Dem. Rep, Congo Rep., Gabon, Russia, Togo, Ukraine) abolished or liberalized their export marketing board through 2010. As neither of these countries have tariffs over 40 percent or black market premiums over 20 percent, they would be considered “open” when liberalizing their export marketing board. Internal conflict; UCDP-PRIO Data from UCDP dataset allows constructing one dummy for internal conflict and war participation Dataset one for war participation. In the latter, we consider a country to be participating in a war only if it is listed either as the country of location, or a major participant (side A or B), omitting countries that are listed as allies. Civil liberties, Freedom Sum of the civil liberties and the political rights indicator, both measured on a 1-7 political rights House scale. http://www.freedomhouse.org/report/freedom-world-2012/methodology Financial Chinn-Ito The Chinn-Ito index (KAOPEN) is an index measuring a country's degree of capital Openness Index account openness. KAOPEN is based on the binary dummy variables that codify the 33 tabulation of restrictions on cross-border financial transactions reported in the IMF's Annual Report on Exchange Arrangements and Exchange Restrictions (AREAER). http://web.pdx.edu/~ito/Chinn-Ito_website.htm Primary WDI Gross primary school enrollment rates (percent of population) schooling Gini coefficient Barro-Lee The Barro-Lee dataset provides data on the percentage of the population that on educational dataset attained different levels of education: No education (0 years), complete primary (6 attainment years), complete secondary (12 years), and complete tertiary (16 years). For non- complete primary, secondary, or tertiary we assume respectively 3 years, 9 years, and 14 years of schooling. With this information, we can construct a Lorenz curve measuring which percentage of population attained which percentage of total years of schooling. With this information, we construct a Gini coefficient that measures educational inequality analogous to the standard income inequality measure. Agricultural WDI WDI Indicator: NV.AGR.TOTL.ZS, “Agriculture, value added (% of GDP)”. Share in GDPproductivity 34 Figure 1: Weights On Incomes and Individuals Implied by Social Welfare Functions Panel A: Weights on Incomes 0.005 Weight on Income of Percentile j in 0.004 Social Welfare Function 0.003 Bottom40 Sen 0.002 Atkinson(0) Atkinson(1) 0.001 Atkinson (2) 0 0 0.2 0.4 0.6 0.8 1 Percentile of Income Distribution Panel B: Weights on Individuals 0.005 Weight on Percentile j in Social Welfare 0.0045 0.004 0.0035 0.003 Bottom40 Function 0.0025 Sen 0.002 Atkinson(0) 0.0015 Atkinson(1) 0.001 Atkinson (2) 0.0005 0 0 0.2 0.4 0.6 0.8 1 Percentiles of Income Distribution Notes: The top panel shows the weights (on the vertical axis) assigned by the indicated social welfare functions to the incomes of individuals at different points in the income distribution (on the horizontal axis). The bottom panel reports the same information, but for the weights assigned to individuals rather than incomes. Weights are normalized to sum to one. Weights are in general data-dependent and are depicted for a hypothetical lognormal income distribution with a mean of $2000 per annum and a Gini coefficient equal to 30. 35 Figure 2: Availability of Household Survey Data (POVCALNET and LIS) 60 50 40 30 20 10 0 1967 1969 1971 1973 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 LIS Database (# obs) PCN Database (#obs) Notes: This figure shows number of household surveys available in each year, for the LIS and POVCALNET databases. 36 Figure 3: Growth in Average Incomes and Growth in Social Welfare Bottom 10% Bottom 20% Bottom 40% Bottom 90% Figure continues on following page 37 Figure , Cont’d: Growth in Average Incomes and Growth in Social Welfare Atkinson Index A(1) Atkinson Index A(2) Atkinson Index A(3) Sen Index Notes: This graph plots the average annual growth rate in average income (on the horizontal axis) against average annual growth rate in social welfare (on the vertical axis), for each indicated social welfare function. All growth rates are constant price annual average log-differences. The dataset consists of the sample of 285 non-overlapping spells lasting at least five years. 38 Figure 4: Distribution of Growth in Average Incomes and Growth in Inequality Bottom 10, 20, 40 and 90 Percent Atkinson A(1) A(2) A(3) and Sen Indices Notes: This graph shows the kernel density estimate of the distribution of the growth rate of average income, and the growth rate of the inequality change component of the indicated social welfare functions. The dataset consists of the sample of 284 non-overlapping spells lasting at least five years. 39 Figure 5: Standard Deviation of Growth Rate of Income Shares Actual (Decile Shares, Sample of Min-5 Year Spells) Standard Deviation of Growth Rate of 0.06 0.05 0.04 Decile Share 0.03 0.02 0.01 0 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 Artificial (Sampling Variation Only, No Changes in True Decile Shares) 0.06 Standard Deviation of Growth Rate of 0.05 0.04 Percentile Share 0.03 0.02 0.01 0 0 0.2 0.4 0.6 0.8 1 Percentile p Notes: The top panel reports the standard deviation (across spells) of the average annual growth rate of the indicated decile share, in the sample of spells at least five years long. The bottom panel graphs the hypothetical standard deviation of the growth rate of the income share of each percentile of the income distribution under the assumption that the only source of variation across spells is sampling variation, i.e. the population mean and variance of log income are the same at both endpoints of each spell. 40 Table 1: Descriptive Statistics Average Bottom Bottom Bottom Bottom Atkinson Atkinson Atkinson Social Welfare Function Sen Index incomes 10% 20% 40% 90% Index (1) Index (2) Index (3) All Observations (Pooled) Sample 1: Long Spells (N=117) Mean 0.020 0.006 0.005 0.003 0.001 0.002 0.003 0.004 0.001 Std. Dev. 0.028 0.033 0.025 0.018 0.008 0.009 0.018 0.024 0.010 P10 -0.011 -0.022 -0.018 -0.015 -0.008 -0.006 -0.012 -0.016 -0.010 P90 0.052 0.042 0.034 0.025 0.012 0.013 0.025 0.031 0.014 Sample 2: Min-5-year Spells (N=285) Mean 0.015 0.002 0.001 0.000 0.000 0.000 0.001 0.002 0.000 Std. Dev. 0.042 0.052 0.034 0.023 0.010 0.012 0.028 0.040 0.012 P10 -0.028 -0.053 -0.039 -0.027 -0.012 -0.014 -0.028 -0.038 -0.015 P90 0.063 0.061 0.045 0.031 0.012 0.016 0.031 0.041 0.016 By Income Category (Min-5-year Spells) Low Income (N=42) Mean 0.025 0.010 0.007 0.004 0.001 0.002 0.005 0.007 0.002 Std. Dev. 0.035 0.051 0.040 0.029 0.015 0.016 0.029 0.038 0.017 P10 -0.012 -0.037 -0.027 -0.028 -0.016 -0.015 -0.022 -0.025 -0.018 P90 0.058 0.076 0.061 0.039 0.023 0.021 0.036 0.055 0.025 Middle Income (N=168) Mean 0.013 0.003 0.002 0.001 0.000 0.001 0.002 0.003 0.000 Std. Dev. 0.049 0.057 0.036 0.024 0.010 0.013 0.032 0.046 0.013 P10 -0.037 -0.060 -0.047 -0.034 -0.013 -0.016 -0.030 -0.042 -0.017 P90 0.079 0.064 0.047 0.034 0.014 0.017 0.033 0.048 0.018 High Income (N=75) Mean 0.016 -0.004 -0.004 -0.004 -0.002 -0.001 -0.002 -0.003 -0.002 Std. Dev. 0.026 0.037 0.021 0.014 0.006 0.006 0.014 0.022 0.008 P10 -0.013 -0.041 -0.031 -0.022 -0.009 -0.009 -0.016 -0.025 -0.012 P90 0.051 0.036 0.022 0.012 0.004 0.006 0.015 0.023 0.006 By Decade (Min-5-year Spells) 1980s (N = 80) Mean 0.002 -0.002 -0.001 -0.002 -0.001 -0.001 -0.001 -0.002 -0.001 Std. Dev. 0.045 0.037 0.026 0.019 0.010 0.009 0.018 0.025 0.011 P10 -0.058 -0.045 -0.035 -0.022 -0.013 -0.011 -0.018 -0.034 -0.015 P90 0.048 0.034 0.024 0.017 0.009 0.008 0.015 0.022 0.010 1990s (N = 198) Mean 0.008 -0.006 -0.004 -0.003 -0.001 -0.001 -0.003 -0.004 -0.002 Std. Dev. 0.044 0.054 0.035 0.024 0.011 0.013 0.029 0.041 0.013 P10 -0.037 -0.073 -0.051 -0.034 -0.014 -0.016 -0.031 -0.045 -0.018 P90 0.055 0.060 0.043 0.033 0.015 0.017 0.031 0.038 0.018 2000s (N=167) Mean 0.029 0.012 0.007 0.004 0.001 0.002 0.006 0.010 0.001 Std. Dev. 0.036 0.052 0.034 0.023 0.010 0.012 0.029 0.042 0.012 P10 -0.014 -0.030 -0.024 -0.021 -0.010 -0.010 -0.018 -0.023 -0.013 P90 0.080 0.075 0.053 0.034 0.012 0.018 0.039 0.057 0.017 th th Notes: This table reports the mean, standard deviation, and 10 and 90 percentiles of the distribution of growth rates of average income (the first column) and the relevant inequality measure corresponding to the indicated social welfare function (the remaining eight columns). The top panel pools all countries and periods. The middle panel distinguishes countries by their income level. The bottom panel reports results by decades. In the bottom panel, each subsample consists of all spells with at least one end-point in the indicated decade. Spells are weighted by the fraction of years of the spell falling in the indicated decade. The dataset in the second and third panels consists of the sample of 284 non-overlapping spells at least five years long. 41 Table 2: Regressions of Social Welfare Growth on Average Income Growth (1) (2) (3) (4) (5) (6) (7) (8) Atkinson Atkinson Atkinson Social Welfare Function Bottom 10% Bottom 20% Bottom 40% Bottom 90% Sen Index Index (1) Index (2) Index (3) Panel A: Sample with spells of minimum 5 year length Avg. growth - Min 5 year spells 1.151*** 1.075*** 1.021*** 0.991*** 1.008*** 1.043*** 1.083*** 1.003*** (0.0842) (0.0626) (0.0467) (0.0223) (0.0238) (0.0468) (0.0640) (0.0273) Number of Observations 285 285 285 285 285 285 285 285 R-squared 0.476 0.650 0.783 0.944 0.925 0.717 0.571 0.921 Variance decomposition 0.413 0.605 0.767 0.952 0.918 0.687 0.527 0.918 P-val. wald test, slope=1 0.0765 0.234 0.654 0.687 0.749 0.362 0.197 0.903 Panel B: Sample with longest spell per country Avg. growth - Long spells 0.961*** 0.953*** 0.955*** 0.975*** 0.970*** 0.943*** 0.932*** 0.974*** (0.162) (0.119) (0.0785) (0.0249) (0.0382) (0.0835) (0.118) (0.0376) Number of Observations 117 117 117 117 117 117 117 117 R-squared 0.403 0.535 0.685 0.927 0.897 0.688 0.538 0.887 Share of variance due to growth 0.420 0.561 0.718 0.951 0.924 0.730 0.578 0.911 P-val. wald test, slope=1 0.810 0.696 0.569 0.324 0.435 0.496 0.565 0.497 Notes: This table reports the results from a regression of growth in the indicated social welfare function (in each column) on growth in average incomes, in the sample of non- overlapping spells lasting at least five years (Panel A), and in the sample of long spells with one spell per country (Panel B). Heteroskedasticity-consistent standard errors reported in parentheses. * (**) (***) indicate significant differences from zero at the 10 (5) (1) percent level. The last row in each panel reports the p-value associated with the test of the null hypothesis that the estimated slope coefficient is equal to one. 42 Table 3: Regressions of Social Welfare Growth on Average Income Growth: Min 5-Year Spells Disaggregated By Income Level (1) (2) (3) (4) (5) (6) (7) (8) Atkinson Atkinson Atkinson Social Welfare Function Bottom 10% Bottom 20% Bottom 40% Bottom 90% Sen Index Index (1) Index (2) Index (3) Panel A: High Income Countries Avg. growth 1.253*** 1.030*** 0.950*** 0.947*** 0.977*** 1.021*** 1.102*** 0.953*** (0.255) (0.156) (0.106) (0.0469) (0.0504) (0.105) (0.163) (0.0635) Number of Observations 75 75 75 75 75 75 75 75 R-squared 0.441 0.600 0.755 0.947 0.939 0.776 0.615 0.904 Variance decomposition 0.352 0.583 0.795 1.000 0.961 0.761 0.558 0.948 P-val. wald test, slope=1 0.329 0.850 0.644 0.268 0.659 0.846 0.536 0.468 Panel B: Middle Income Countries Avg. growth 1.164*** 1.091*** 1.033*** 0.997*** 1.013*** 1.055*** 1.097*** 1.011*** (0.0955) (0.0719) (0.0544) (0.0264) (0.0279) (0.0544) (0.0737) (0.0319) Number of Observations 168 168 168 168 168 168 168 168 R-squared 0.509 0.693 0.819 0.958 0.936 0.728 0.583 0.939 Share of variance due to growth 0.438 0.636 0.793 0.960 0.924 0.690 0.531 0.929 P-val. wald test, slope=1 0.0911 0.211 0.551 0.922 0.630 0.317 0.193 0.728 Panel C: Low Income Countries Avg. growth 0.899*** 0.951*** 0.967*** 0.975*** 0.978*** 0.947*** 0.923*** 0.975*** (0.291) (0.216) (0.145) (0.0535) (0.0742) (0.159) (0.223) (0.0722) Number of Observations 42 42 42 42 42 42 42 42 R-squared 0.269 0.409 0.569 0.836 0.820 0.569 0.418 0.801 Share of variance due to growth 0.299 0.430 0.588 0.858 0.839 0.600 0.453 0.821 P-val. wald test, slope=1 0.732 0.823 0.821 0.650 0.769 0.744 0.732 0.733 Notes: This table reports the results from a regression of growth in the indicated social welfare function (in each column) on growth in average incomes, in the sample of non-overlapping spells lasting at least five years. The three panels correspond to sets of countries at different income levels, following the World Bank’s classification for 2012. Heteroskedasticity-consistent standard errors reported in parentheses. * (**) (***) indicate significant differences from zero at the 10 (5) (1) percent level. The last row in each panel reports the p-value associated with the test of the null hypothesis that the estimated slope coefficient is equal to one. 43 Table 4: Regressions of Social Welfare Growth on Average Income Growth: Min 5-Year Spells Disaggregated By Decade (1) (2) (3) (4) (5) (6) (7) (8) Social Welfare Function Bottom 10% Bottom 20% Bottom 40% Bottom 90% Atkinson Index (1) Atkinson Index (2) Atkinson Index (3) Sen Index Panel A: 1980s Avg. growth 1.120*** 1.061*** 1.016*** 0.980*** 0.995*** 1.019*** 1.049*** 0.998*** (0.133) (0.0991) (0.0753) (0.0388) (0.0375) (0.0689) (0.0941) (0.0468) Number of Observations 80 80 80 80 80 80 80 80 R-squared 0.664 0.782 0.859 0.956 0.959 0.870 0.780 0.943 Variance decomposition 0.593 0.737 0.846 0.975 0.964 0.854 0.743 0.945 P-val. wald test, slope=1 0.369 0.542 0.835 0.616 0.904 0.785 0.604 0.961 Panel B: 1990s Avg. growth 1.127*** 1.062*** 1.013*** 0.983*** 0.999*** 1.025*** 1.058*** 0.998*** (0.0979) (0.0757) (0.0590) (0.0297) (0.0306) (0.0559) (0.0739) (0.0349) Number of Observations 198 198 198 198 198 198 198 198 R-squared 0.468 0.650 0.779 0.940 0.924 0.716 0.569 0.919 Share of variance due to growth 0.415 0.612 0.769 0.956 0.924 0.699 0.538 0.921 P-val. wald test, slope=1 0.196 0.417 0.824 0.579 0.977 0.662 0.436 0.950 Panel C: 2000s Avg. growth 1.018*** 0.996*** 0.970*** 0.985*** 0.990*** 0.991*** 1.000*** 0.982*** (0.119) (0.0776) (0.0505) (0.0189) (0.0261) (0.0638) (0.0935) (0.0260) Number of Observations 167 167 167 167 167 167 167 167 R-squared 0.326 0.517 0.699 0.931 0.890 0.589 0.416 0.895 Share of variance due to growth 0.320 0.519 0.720 0.945 0.899 0.595 0.416 0.911 P-val. wald test, slope=1 0.879 0.959 0.557 0.444 0.709 0.887 0.996 0.481 Notes: This table reports the results from a regression of growth in the indicated social welfare function (in each column) on growth in average incomes, in the sample of non- overlapping spells lasting at least five years. The three panels correspond to the indicated time periods. The regression for each decade includes all spells with at least one end- point falling in the indicated decade. Spells are weighted by the fraction of the years of the spell falling in the indicated decade. Heteroskedasticity-consistent standard errors reported in parentheses. * (**) (***) indicate significant differences from zero at the 10 (5) (1) percent level. The last row in each panel reports the p-value associated with the test of the null hypothesis that the estimated slope coefficient is equal to one. 44 Table 5: Regressions of Social Welfare Growth on Average Income Growth: Alternative Measures of Real Income Growth Different Welfare Measures (1) (2) (3) (4) (5) (6) (7) (8) Atkinson Index Atkinson Index Atkinson Index Bottom 10% Bottom 20% Bottom 40% Bottom 90% Sen Index (1) (2) (3) Panel A: Survey data (income or consumption) Avg. growth - Min 5 year spells 1.075*** 1.006*** 0.965*** 0.963*** 0.983*** 1.005*** 1.038*** 0.967*** (0.100) (0.0695) (0.0501) (0.0238) (0.0275) (0.0572) (0.0798) (0.0283) Number of Observations 274 274 274 274 274 274 274 274 R-squared 0.383 0.565 0.722 0.927 0.902 0.645 0.485 0.899 Share of variance due to growth 0.356 0.561 0.749 0.963 0.917 0.642 0.467 0.929 P-value of wald test, slope=1 0.454 0.930 0.481 0.123 0.545 0.934 0.638 0.254 Panel B: Real private consumption per capita (National accounts data) Avg. growth - Min 5 year spells 1.146*** 1.051*** 1.011*** 0.997*** 1.011*** 1.060*** 1.105*** 1.002*** (0.0938) (0.0648) (0.0463) (0.0212) (0.0241) (0.0513) (0.0723) (0.0259) Number of Observations 274 274 274 274 274 274 274 274 R-squared 0.301 0.463 0.634 0.891 0.855 0.552 0.394 0.851 Share of variance due to growth 0.262 0.440 0.626 0.894 0.846 0.521 0.357 0.849 P-value of wald test, slope=1 0.122 0.433 0.805 0.882 0.660 0.241 0.148 0.925 Panel C: Mixed measure (arithmetic average from A&B) Avg. growth - Min 5 year spells 1.133*** 1.030*** 0.977*** 0.969*** 0.992*** 1.034*** 1.082*** 0.975*** (0.109) (0.0768) (0.0551) (0.0262) (0.0299) (0.0607) (0.0846) (0.0312) Number of Observations 274 274 274 274 274 274 274 274 R-squared 0.299 0.457 0.621 0.888 0.852 0.543 0.388 0.847 Share of variance due to growth 0.264 0.443 0.636 0.916 0.859 0.526 0.358 0.869 P-value of wald test, slope=1 0.226 0.696 0.677 0.232 0.782 0.581 0.334 0.423 Notes: This table reports the results from a regression of growth in the indicated social welfare function (in each column) on growth in average incomes, in the sample of non-overlapping spells lasting at least five years. The three panels correspond to the three different measures of mean income. Heteroskedasticity- consistent standard errors reported in parentheses. * (**) (***) indicate significant differences from zero at the 10 (5) (1) percent level. The last row in each panel reports the p-value associated with the test of the null hypothesis that the estimated slope coefficient is equal to one. 45 Table 6: Welfare Comparisons Using Generalized Lorenz Dominance Positive Growth Spells Negative Growth Spells All Spells Proportion for which welfare Proportion for which welfare Proportion for which welfare Number of spells Number of spells Number of spells unambiguously increased unambiguously decreased unambiguously changed Min 5 year spells 203 81% 82 65% 284 76% Longest Spell per country 97 85% 20 57% 117 79% Notes: This table summarizes the results of a generalized Lorenz dominance comparison of the initial and final income distributions of each spell in the indicated sample. The first two columns consider spells with positive growth in the survey mean and report the fraction of cases for which the end-point distribution of the spell generalized Lorenz- dominates the initial distribution. The second two columns do the reverse for spells with negative growth in average incomes. The last two columns combine information for all spells. 46 Table 7: Policies and Social Welfare Growth Panel A: SWF is Bottom 40 Percent Panel B: SWF is Atkinson Index (Theta=3) Panel C: SWF is Sen Index Posterior Posterior Posterior Inclusion Social Inclusion Social Inclusion Social Probability Welfare Growth Equality Probability Welfare Growth Equality Probability Welfare Growth Equality Initial income level 1.000 -2.835 -3.055 0.219 1.000 -2.832 -3.015 0.184 1.000 -2.900 -3.081 0.181 (100) (100) (0) (98) (100) (0) (100) (100) (1) Initial equality level 1.000 -1.011 0.078 -1.089 1.000 -2.042 -0.005 -2.037 1.000 -0.496 0.118 -0.615 (94) (4) (100) (100) (0) (100) (14) (8) (100) Financial depth (M2 % GDP) 0.003 -0.001 -0.002 0.001 0.004 0.000 -0.002 0.002 0.002 -0.001 -0.001 0.000 (0) (2) (11) (0) (2) (18) (0) (2) (20) Inflation rate 1.000 -1.980 -1.692 -0.288 1.000 -2.242 -1.705 -0.537 1.000 -1.841 -1.694 -0.147 (100) (100) (12) (99) (100) (20) (100) (100) (7) Budget Balance 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 (23) (44) (0) (19) (41) (0) (31) (47) (0) Trade Openness 0.004 0.001 0.002 -0.001 0.004 0.001 0.002 -0.001 0.008 0.003 0.003 -0.001 (21) (26) (0) (13) (26) (0) (25) (26) (0) Population growth 0.575 -0.508 -0.394 -0.114 0.762 -0.904 -0.519 -0.385 0.563 -0.452 -0.380 -0.071 (51) (10) (48) (95) (14) (88) (24) (8) (66) Life expectancy 0.039 0.006 0.016 -0.010 0.037 0.000 0.015 -0.014 0.048 0.015 0.020 -0.005 (1) (11) (2) (1) (12) (1) (6) (11) (1) Revolutions per pop. 0.470 -0.288 -0.276 -0.012 0.308 -0.226 -0.191 -0.036 0.579 -0.351 -0.331 -0.020 (44) (54) (0) (39) (57) (0) (54) (53) (0) Civil Liberties / Democracy 0.031 -0.008 -0.010 0.003 0.027 -0.001 -0.008 0.008 0.035 -0.010 -0.012 0.002 (0) (0) (0) (0) (0) (0) (0) (0) (0) Internal/external conflict (dummy 0.034 0.002 -0.002 0.005 0.038 0.007 -0.001 0.009 0.033 0.000 -0.002 0.003 (1) (0) (0) (1) (0) (0) (0) (0) (0) Fin. openness (Chinn-Ito) 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 (13) (9) (0) (24) (10) (0) (12) (9) (0) Primary school enrollment rate 0.003 0.000 0.001 -0.001 0.004 -0.002 0.001 -0.003 0.008 0.002 0.003 -0.001 (3) (2) (68) (33) (1) (97) (0) (2) (50) Education Gini 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 (17) (34) (0) (14) (35) (0) (25) (31) (0) Agriculture (% GDP) 0.997 -1.969 -2.265 0.296 0.969 -1.701 -2.197 0.496 0.999 -2.090 -2.260 0.171 (88) (100) (0) (51) (100) (0) (100) (100) (0) Notes: This table summarizes selected BMA results. The three panels correspond to specifications using the three indicated social welfare functions. Within each panel, the first column reports the posterior inclusion probability for the indicated explanatory variable. The second column within each panel reports the posterior probability- weighted estimated slope coefficients for the growth rate of social welfare (Equation (9)). The third and fourth columns within each panel report the corresponding posterior probability-weighted slopes for the regressions where growth in average incomes, and growth in the relevant equality measure, are the dependent variable. As discussed in the text, these two slope coefficients sum to the slope coefficient reported in the second column, and they are multiplied by the standard deviation of the corresponding right-hand-side variable. The number in parentheses indicate the fraction of models in which the estimated slope coefficient is statistically significantly different from zero and of the same sign as the posterior probability-weighted slope. 47