Trade and Inequality in the US and the UK: A Cohort Perspective. Ozan Eksi. TOBB University of Economics and Technology. June 2010

Transcription

1 Trade and Inequaliy in he US and he UK: A Cohor Perspecive Ozan Eksi TOBB Universiy of Economics and Technology June 2010 Absrac: This sudy examines how aggregae inequaliy indices give long-lasing responses o srucural changes in he economy because of he cohor srucure of he populaion. This long-lasing propery of response may lead o he failure o deec relaionships beween inequaliy and oher variables in esimaes ha measure linear correlaion, including regressions. An alernaive measure of inequaliy, he average growh rae of wihin-cohor inequaliies, finds ha he relaionship beween rade and inequaliy is posiive and significan. This resul, which is consisen wih rade heory and a subsanial par of he rade lieraure, conflics wih evidence obained by aggregae inequaliy indices. JEL Codes: D31, F40 Keywords: Inequaliy, Inernaional Trade Considerable research and lieraure has addressed he effec of rade liberalizaion on increasing inequaliy across developed counries, and a subsanial par of his lieraure explains his effec hrough he rising skill premium, based on predicions of rade heory. According o Solper-Samuelson (1941), rade globalizaion should increase wihin-counry earning inequaliies in he developed regions of he world because of he increase in demand for goods produced by skilled labor (Norh and Souh rade), which increases wage dispersion among skilled and unskilled labor groups. However, he empirical inequaliy lieraure ha invesigaes rade in oupus finds lile suppor for his heory. Mos of hese sudies use US daa, and alhough some sudies find a significan effec (Wood 1994, Leamer 1996), mos conclude ha rade volume is no sufficien o explain much of he increase in inequaliy (Lawrence and Slaugher 1993, Sachs and Shaz 1996, Kaz and Murphy 1992). More recen lieraure also considers inermediae inpu rade, which is heoreically moivaed by eiher increasing reurn o scale in he producion of inermediae inpus or by he increasing variey of inpus, which increase he amoun of he final good produced (Krugman 1979, 80). Mos of his ype of rade The firs draf of his paper was par of my disseraion compleed a he Universia Pompeu Fabra. I am graeful o Thijs van Rens and o Vasco Carvalho, Jaume Venura, and Sergi Jiménez from ha insiuion; o Chris Telmer from Carnegie Mellon Universiy; and o John Weeks from he Universiy of London. I also hank o my curren colleagues a TOBB Universiy of Economics and Technology for heir useful commens and suggesions. 1

2 akes place beween developed counries (Miroudo e al. 2009). Even hough he heory using inpu rade is more successful han he heory using oupu rade in explaining he rise in earning inequaliy, i sill does no accoun for much of his rise (Feensra and Gordon 1995, 1996). Trade lieraure ends o indicae ha skill-biased echnological change has been a more imporan deerminan of inequaliy han is rade iself (Acemoglu 2000, Krugman 2000). The skill premium explanaion is only one dimension of wage inequaliy. Once we divide he labor force ino skilled and unskilled labor, we mus also consider ha wage inequaliy could change because of changes in wihin group variances. Boh Juhn, Murphy and Pierce (1993) and Goschalk (1997) documen ha inequaliy increased no only beween groups caegorized by experience and educaion, bu also wihin hese groups. Goschalk (1997) furher documens ha experience and educaion are equally imporan o he rise of earning inequaliy. However, o my knowledge, no sudy has relaed wihin-group earning inequaliies o rade. Trade can also affec earning inequaliy by changing he relaive share of skilled and unskilled labor in he labor force (e.g., by shifing he labor force away from he agriculure secor). There are also facors hrough which rade may affec income bu no wage inequaliy. Comin and Philippon (2006) address he increase in he urnover of marke shares from increasing compeiion in he produc marke, and Di Giovanni and Levchenko (2006) documen ha he producion of secors ha are open o inernaional rade are less correlaed wih he res of he economy han are oher secors. This laer finding is inuiive, as raders are subjec o foreign shocks. This decline in he correlaion of shocks across secors is expeced o increase income inequaliy by dispersing income. Unlike he large amoun of lieraure ha aemps, wih lile success, o explain he effec of rade on beween-group wage inequaliy (skill premium), he lieraure regarding he oher channels hrough which rade can affec inequaliy is scarce, bu wha here is of i ends o indicae ha rade increases inequaliy. In empirically examining he effec of rade on inequaliy, aking ino accoun all possible channels, one may refer o regression esimaes ha use aggregae inequaliy measures. These esimaes find weak suppor for rade s effec on increasing inequaliy, and some show ha rade is a moderaor of inequaliy across a sample of indusrialized counries (IMF WEO, 2007). The presen sudy presens conrary evidence ha rade is an imporan deerminan of inequaliy. Firs, he sudy provides graphical evidence showing ha changes in inequaliy occur a imes of increasing rade (and financial) liberalizaion across a sample of developed counries. Second, i explains he lack of suppor for he variables relaionship found in ime series regression esimaes based on a finding in Eksi (2010). In ha sudy I discuss how aggregae inequaliy indices give a long-lasing responses o changes in he srucure of he economy, so long as hese changes affec he income pa- 2

3 ern of individuals. I claim ha, when income paerns change, he old cohors have already been affeced by heir previous incomes so he change in inequaliy should las unil he old cohors leave he populaion, replaced by young cohors. Consequenly, regression esimaes ha measure insananeous correlaion beween rade and inequaliy are inappropriae. The presen sudy analyzes his issue furher and suggess an alernaive measure of inequaliy o be used in regression esimaes: he growh rae of wihincohor inequaliies. This sudy begins wih an analysis of inequaliy measures and derives he difference beween growh in aggregae (average wihin-cohor) inequaliy and he average growh rae of wihin-cohor inequaliies. This analysis shows ha he measure of inequaliy ha should be used in regression esimaes is he average growh rae of wihincohor inequaliies. Then he derivaions are aken o he daa in order o revisi he effec of rade on increasing inequaliies in he US and he UK. These counries are chosen based on he availabiliy of he high-qualiy micro income daa in he form of crosssecional daa ses ha allows consrucing synheic cohor daa o measure wihincohor inequaliies. The esimaion begins by replicaing IMF WEO (2007) esimaes, which explain he change in he Gini coefficien across a sample of counries by using a se of variables ha includes rade and financial globalizaion indicaors, educaion level, financial and echnological developmen indicaors, and secoral employmen shares. I firs use he aggregae Gini index in order o replicae he IMF resuls for he US and he UK, and hen use a panel of wihin-cohor Gini indices from hese counries in he same regression. The resuls found by aggregae indices are consisen wih IMF esimaes: rade has decreased inequaliy in hese counries, alhough he esimaed coefficiens are no significan. The resuls obained wih he growh rae of cohor inequaliies which have heory consisen signs for he oher explanaory variables as well indicae ha rade had a posiive and significan effec on inequaliy in he US and he UK. These resuls are robus o changes in conrol variables, sample period, and changes in dummy variables used in panel esimaes. Using wihin-cohor dispersion of incomes is no a perfec measure of inequaliy because people in he same cohor age simulaneously, so hey have he same level of experience hroughou heir life spans. Hence, inequaliy canno be expeced o increase because of changes in reurn o experience. However, reurn o experience is par of he skill premium, and he lieraure has already shown ha he effec of rade on he skill premium is significan. As a resul, compared o he cos of using he aggregae inequaliy index, he cos of using wihin-cohor inequaliy daa mus be much smaller. The res of he paper is organized as follows. Secion 1 reviews he inequaliy measures and explains he difference beween growh in aggregae (average wihin- 3

4 cohor) inequaliy and average growh in wihin-cohor inequaliy. Findings are suppored by a simulaion sudy as well. Secion 2 discusses why rade is relevan in explaining he recen increases in inequaliy in he US and he UK. Secion 3 presens he framework for he empirical analysis, while secion 4 explains he daa and income inequaliy paerns in he US and he UK. Secion 5 compares differen measures of inequaliy on he real daa, and secion 6 presens he empirical resuls on he effec of rade on inequaliy obained wih hese measures. Secion 7 concludes. 1 Growh in Average Wihin-Cohor (Aggregae) Inequaliy vs. Average Growh in Wihin-Cohor Inequaliies This secion explains ha he regression esimaes using aggregae inequaliy measures (e.g., he Gini coefficien, he variance of log incomes) are subjec o problems ha sem from he ime effec. Anoher measure of inequaliy is proposed for use in regression esimaes: he average growh in wihin-cohor inequaliies. 1.1 Individual Income Process A reduced form income process describes a pah for log real income y for N individuals i over ime periods : yi α yi 1 υ i (1) Individuals' incomes are subjec o shocks wih a mean of zero across individuals and ime, E ( υ ) 0; have he same variance in ime across individuals, i i are correlaed among individuals, Corr( υ, υ ) i j subscrip, as hey may no be consan over ime. 1 Var ( υ ) 2 s i σ ; and ρ. Second momens are denoed wih a 1.2 Dynamics of Aggregae Inequaliy Eksi (2010) discusses how variance and cross-secional correlaion characerisics of he income shocks, υ i, can affec aggregae and wihin-cohor inequaliies. Wihin-cohor inequaliy measures he inequaliy wihin a group of individuals born in he same year who have been subjec o income shocks differen from hose experienced by each oher individual in he group over ime. These shocks, even if heir variance and correlaion are consan over ime, disperse income and increase wihin-cohor inequaliy. This effec is called he "age effec" on inequaliy. However, even hough wihin-cohor inequaliy increases over ime, as long as he rae of is increase is consan, wihin-counry (aggregae) income inequaliy says a he same level over ime because old cohors are 1 In equaion (1) he error erm has a permanen effec on income. Permanen shocks are chosen o sar wih since he inuiion for he mechanism sressed in his paper is bes undersood in a model wih only his ype of shock. 4

5 replaced by younger ones, so similar cohors make up he populaion a each poin in ime. The dynamics of wihin cohor and aggregae inequaliies are as shown in Figure 1. 2 Relaive Incomes Cohor 3,-1 Cohor 2,-1 Cohor 3, Cohor 1,-1 Cohor 2, Cohor 1, I () = I(-1) Time Figure 1. Illusraion of he Evoluion of Income Disribuion of Three Cohors Noes: The figure shows he age effec in shaping income disribuion using hree cohors, each wih a life span of hree years. As an example, Cohor 2, represens people who are wo years old a ime. The age effec excludes he populaion effec, so he birh rae is consan over ime; he cohor effec, so he shocks have equal disribuion among cohors; and he ime effec, so cohors are unaffeced by unexpeced changes in income disribuion ha occur a a poin in ime, such as hose ha can occur wih a ax reform or a change in a social securiy sysem. Figure 1 indicaes ha he aggregae inequaliy a ime is equal o he aggregae inequaliy a ime -1, despie he increase in wihin-cohor inequaliies. Figure 2 shows he effec of an increase in he variance of income shocks. Relaive Incomes Cohor 3,-1 Cohor 2,-1 Cohor 1, Time Figure 2. The Effec on Equaliy of an Increase in Variance (ime effec) Figure 2 shows ha, following an increase in variance, increases in boh wihincohor and aggregae inequaliies are observed beween imes -1 and. A similar figure would be aained if we had assumed a decline in he correlaion of he income shocks. The figure also shows ha, a he ime of a decline in variance, he old cohors 2 Alhough Figure 1 uses discree ime for noaion, he lines and he income disribuion are drawn consisen wih coninuous ime. The normally disribued incomes are aained because of he higher number of individuals around he mean income. 5

6 have already been affeced by he previous shocks, so we can infer ha only a ime +2 have all he cohors compleely adaped o he decline in variance so he aggregae inequaliy can reach is new seady sae. 3 Figures 1 and 2 also indicae ha he difference in individuals income growh rae, α i, canno lead o a change in aggregae inequaliy, as ha growh rae is consan over ime, so a difference in income growh rae would explain only par of inequaliy. However, ime-varying parameers aside, some classificaions of he facors ha can affec aggregae income inequaliy are more general: populaion effec, cohor, and ime effecs explained in he noe under Figure 1.Because indusrialized counries make up he sample in his paper, none of hese condiions is expeced o have a significan effec on he resuls. Eksi (2010) shows ha i ook around a decade for he US income inequaliy o adjus o he changes in he US economy during he Grea Moderaion. This secion considers an implicaion of cohor srucure of he populaion on he measures of inequaliy and re-defines an individual income process ha is also specified for each cohor. Consider an economy ha is populaed by individuals i of differen ages a. An individual of age a a ime belongs o he cohor c=-a, where cohors are defined by birh year. The log income of each individual y i, c, evolves according o yi, c, αi yi, c, 1 ν i,, (2) where he shock ν i, has he same characerisics as ha in equaion (1). However, unlike equaion (1), equaion (2) uses an idiosyncraic rend for he long-erm growh rae i of income, α, wih he assumpion ha his idiosyncraic par is uncorrelaed wih everyhing else and ha i has a consan variance. Figure 2 suggess ha, as long as he disribuion of σ is consan, i affecs he evoluion of wihin-cohor income inequaliy α bu no ha of aggregae inequaliy, which is why i is no needed in equaion (1). The derivaions use he variance of log incomes as a measure of income inequaliy. 4 Wihin-cohor inequaliy a ime is defined as 2 σ α D Var( y ) c, i,, α Calculaing he change in his variable over ime (derivaions are in Appendix A): Dc, Dc, 1 σα (1 ρ ) σ (3) 3 I call a change in ρ or σ occurring a ime he ime effec. However, even if his change remains from ime on such ha i is no specific o ime, i sill affecs only he ime inervals ha are (Guvenen 2009). 4 Van Rens and Teulings (2008) show ha he Gini index and he variance of log incomes measures are monoonic ransformaions of each oher and ha mapping among he variables is possible. 6

7 Equaion (3) indicaes ha he changes in σ and ρ affec inequaliy hrough he 2 2 erm (1 ρ ) σ, and he correc measure for his erm is Dc, Dc, 1 σ α. To explore wha caused he changes in hese parameers, Dc, is regressed on candidaes X. Dc, β0 β1 X u c,, 2 2 where we would expec β0 σα and β1 d[(1 ρ ) σ ]/ dx. Clearly, his regression is equivalen (in erms of poin esimaes, no sandard errors) o ( D ) β β X ( u ). (4) c, 0 1 c, As a resul, we can also use he average change in he wihin-cohor inequaliies. Wha would happen if we use he change in average wihin-cohor inequaliy as in equaion (5), raher han he average change in wihin-cohor inequaliy? ( D ) γ γ X ν (5) c, 0 1 If he populaion size is consan, cohors have equal size, and if shocks are uncorrelaed across cohors, aggregae inequaliy equals average wihin-cohor inequaliy. Using his insigh, assume ha a new cohor is born in each period and ha everyone in he cohor dies A years laer such ha, a any poin in ime, here will be A cohors alive since he cohor born in period dies in period A 1. Assume furher ha all cohors have he same level of inequaliy a birh, denoed by D 0. Ieraing he expression for D backward unil birh gives c, Dc, Dc, 1 σα (1 ρ ) σ D 2 σ (1 ρ ) σ (1 ρ ) σ 2 c, 2 α a α (1 a s ) a s s 1 D aσ ρ σ (6) Then he average inequaliy among he cohors living a ime is: 1 A 1 a ( Dc, ) ( D0 aσα (1 ρ a s ) σ a s ) A a 0 s 1 The change in he erm wrien in he previous line can be found as 1 ( D ) [(1 ρ ) σ (1 ρ ) σ ] A 1 a c, a s a s 1 a s 1 a s A a 0 s 1 For simpliciy, suppose ha here is a srucural change in such ha (1 ρ ) σ (1 ρ ) σ when 0 0 (1 ρ ) σ (1 ρ ) σ when 1 1 ρ and 2 σ in period, Then he growh rae of he average cohor inequaliy reduces o 7

8 and one period laer o A 1 ( Dc, ) [(1 ρ1 ) σ1 (1 ρ0 ) σ0 ] A A 2 ( Dc, 1) 1 [(1 ρ1 ) σ1 (1 ρ0 ) σ0 ] A where he righ-hand side of he expression dies ou linearly for furher increases in he periods and he changes in ρ and can be inferred from he Figure 2. 5 σ have he greaes effec on inequaliy a ime as From he definiion of D c, in equaion (3), we can also infer ha, upon a srucural change in ρ and σ in period, D c, can be wrien as ( D ) aσ (1 ρ ) σ when c, α 0 0 ( D ) aσ (1 ρ ) σ when c, α 1 1, Hence, ( D ) ( D ), c, c, which indicaes ha using he change in average wihin-cohor inequaliy in he regressions is no he same as using he average of differences in wihin-cohor inequaliies. This analysis suggess ha using wihin-cohor inequaliy daa can help o overcome ime effec he changes in eiher variance or correlaion a cerain poins in ime and follow he rue effec of srucural changes in he economy ha influence individuals income paerns and, as a resul, he income disribuion of he sociey Simulaion The resuls of a simulaion sudy are presened o show how changes in eiher he correlaion or variance of individual income shocks have a long-lasing effec on aggregae inequaliy indices. The sudy uses an income process consisen wih (2) o simulae he income of 30 cohors, each wih 500 unis (households, families, or individuals). Figure 4 shows he resuls of he simulaion sudy. Graph (a) in Figure 4 shows he evoluion of inequaliy in one of he cohors in he US, using real daa. This resul is used o calibrae he wihin-cohor inequaliy daa simulaed, shown on graph (b). On boh graphs inequaliy increases wihin cohors, and he rae of his increase declines over ime. The levels of Gini indices differ as in he simulaion exercise cohors are defined as having no inequaliy when hey ener he labor force; hence, in he simulaed daa he iniial Gini index is 0. Graph (b) is aained when he sandard deviaion of shocks is kep a 0.05, as shown in Graph (c). In his case, we also obain consan aggregae inequaliy, shown in he iniial years of Graph (d). When 5 In a furher aemp o make hese derivaions comprehensible, I give a numerical example by calibraing A o 4 in Appendix B. 8

9 he sandard deviaion of he shocks rises by 25 percen, which one can assume represens he effec of srucural changes in he economy including rade liberalizaion graphs (d), (e), and (f) show he response of differen inequaliy measures. Graph (d) shows ha, wih 30 cohors, i akes 30 years for aggregae inequaliy o finish responding o his increase and o reach he new seady sae. Using he growh rae of aggregae inequaliy, Graph (e) shows ha a change in he variance of he shocks has he greaes effec on aggregae inequaliy in he iniial period of change, afer which he effec dies ou linearly, as prediced in Secion 1.2 and discussed in Appendix B. Finally, graph (f) shows he response of he average growh rae of wihin-cohor inequaliies. Like he shif in variance, here is a shif in his variable, which happens as he rae of growh of wihin-cohor inequaliies changes wih changes in variance or correlaion. Graph (f) is no exacly horizonal, as he change in variance is, as he correlaion could no be oally fixed in he simulaions wih a limied number of households. 2 Why Trade? My conjecure is ha aggregae inequaliy indices give a long lasing response o all ypes of srucural changes in he economy so long as hey affec he income paern of individuals. Hence, I do no expec inernaional rade o be he only variable whose effec on inequaliy can be jusified hrough he growh raes of wihin-cohor inequaliies. However, in Eksi (2010) I calculae he changes in variance and cross-secional correlaion of individuals' income shocks in he US and I propose ha no only an increase in he variance of individual income shocks, bu also a decline in heir correlaion across individuals conribued o he changes in inequaliy and GDP volailiy across indusrialized counries in he ime period considered. GDP volailiy refers o he flucuaions in aggregae income, which can be decomposed o individual income shocks. To deermine wha ype of srucural change could cause his phenomenon, I presen evidence ha demonsraes ha rade may be responsible for he changes in he inequaliy daa. This se of evidence is consisen wih he predicions of he economic heory discussed in he inroducion of his ex bu is conrary o he resuls of regression esimaes obained hrough he sample of indusrialized counries (IMF, WEO 2007). I was his discrepancy ha moivaed me o concenrae on rade. Bu how rade would explain he decline in cross-secional correlaion of income shocks? If we realize ha aggregae shocks ha affec every individual in a parallel manner and, as a resul, creae correlaions beween hem, he decline in correlaion can be explained by a decline in he variance of aggregae shocks, an increase in he variance of idiosyncraic shocks, or he inroducion of a new shock ha decreases he weigh of he aggregae shock in he oal variance of he shocks. Given ha he changes in he daa occurred in previous decades, he las explanaion seems he mos likely candidae if he 9

10 new shock is he foreign shock ha appeared hrough globalizaion. These shocks could be he demand shocks ha affec exporers or he supply shocks ha affec imporers. My predicion is ha, wheher he shock is rade or financial inegraion, i should affec inequaliy he same way: hey are boh forms of opening up a counry o foreign shocks. In his way he incomes of individuals (or heir indusries) who are direcly inegraed wih he res of he world become less correlaed wih he res of he individuals living in he same counry ha is, individuals whose incomes are affeced only by he general equilibrium effecs of inegraion. This inuiion finds suppor in Comin and Philippon (2006) and in Figures 4, 5 and 6, which show ha changes in inequaliy in he sample counries occur when hese counries experience an increase in financial and rade inegraion along wih he res of he world. 6 Wha hese figures show is ha, as soon as an increase in any ype of globalizaion indicaor is observed, an increase in inequaliy occurs hroughou he counries, likely because of a decline in correlaion. Ialy and Denmark may seem o be excepions o his generalizaion, bu a close look a heir figures reveals ha he prediced effec of globalizaion is he same: globalizaion causes a decline in correlaion, which sops an iniial decline in inequaliy a he naional level. Therefore, inequaliy follows a U-shaped paern across he sample counries (which were chosen only based on he availabiliy of inequaliy daa). I focus on rade, raher han financial liberalizaion, because he share of rade globalizaion in he GDP is much greaer han he share of globalizaion wih respec o financial indicaors across developed counries including he UK and he US, making i easier o raionalize is effecs on inequaliy. In addiion, financial inegraion has already been shown o increase inequaliy hrough regression esimaes (IMF 2007). If we accep ha rade liberalizaion may have played a role in a change in aggregae inequaliy, he moivaion for his sudy can be seen in Figure 7, a subse of Figure 4. Figure 7 6 These counries are chosen o demonsrae he relaionship beween inequaliy and rade in a larger sample of counries han he US and he UK. These figures also use he daa I consruced in Eksi (2010). 10

11 In he UK rade increases suddenly up o some consan level, bu i akes a decade for he aggregae inequaliy o respond his change. The daa for he US does no show such a clear paern, bu i reflecs he same idea as well. Hence, Figure 7 could explain why regression esimaes do no find srong evidence of a correlaion beween rade and inequaliy. Using he panel esimaes, I discuss and accoun for possible channels oher han rade, bu I focus on he rade explanaion because i is consisen wih he changes in he micro daa and he volume of rade makes sudden jumps over ime in many developed counries, making i difficul o associae i wih increasing inequaliy, alhough his issue seems o be less serious in he US daa. 3 Framework for he Empirical Analysis Empirical analysis reveals he differences beween esimaes of rade effec on inequaliy when wo differen measures of inequaliy, aggregae inequaliy and he growh rae of wihin-cohor inequaliies in panel regressions, are used as dependen variables. The esimaions are based on he framework of Chaper 4 of he IMF WEO (2007), which explains he change in he Gini coefficien across a sample of counries using a se of variables.. Because of limied availabiliy of high-qualiy micro daa, I use only he US and UK daa in he esimaions. The basic IMF model relaes he Gini index o several measures of globalizaion using a se of conrol variables as shown in Eq. (7): X M A L ln( Gini) α1 α2 ln( ) α3 ln(100 Tariff ) α4 ln( ) Y Y K ICT Credi α5kaopen α6 ln( ) α7 ln( ) α8popshare (7) K Y E AGR EIND α9 ln( H) α10 ln( ) α11 ln( ) ε E E where X and M are non-oil expors and impors, respecively; Y is GDP; TARIFF is he average ariff rae; A and L are cross-border financial asses and liabiliies, respecively (he summaion of A and L is replaced wih he FDI sock in he auxiliary regressions); KAOPEN is he capial accoun openness index; KICT is he raio of informaion communicaion echnology capial sock o he oal capial sock; CREDIT is he volume of credi in he privae secor offered by deposi banks and oher financial insiuions; POPSHARE is he percenage of he populaion aged 15 and older wih secondary or higher educaion; H is average years of educaion in he populaion aged 15 and older; E is oal employmen; and EAGR and EIND are employmen in agriculure and indusry, respecively. The IMF auhors run a panel regression wih fixed effecs (o capure crosscounry differences) and ime dummies. The auhors jusify he use of he logarihm of he Gini index wih heir aim of making his bounded variable behave more like a nor- 11

12 mally disribued variable han is level suggess in order o render i more amenable o OLS esimaion. 7 The model in Eq. (7) deals wih he effec of rade and financial liberalizaion on inequaliy. The World of Work Repor (2008) gives a horough summary of he lieraure on he effec on equaliy of financial liberalizaion; in heory, financial inegraion, helps o reduce wihin-counry income inequaliy by lowering he borrowing consrains of he households wih he leas access o financial markes, bu i can also lead o an increase in inequaliy eiher hrough increased demand for skilled labor or by increasing he likelihood of financial crises, which may disproporionaely hur he poor. Solper- Samuelson (1941) assume inernaional immobiliy of facors of producion, bu FDI weakens his assumpion: Inward FDI ends o ake place in skill- and echnologyinensive secors in he advanced economies, and ouward FDI invess in heir relaively less skill-inensive secors. Empirically, boh inward and ouward FDI are suggesed o increase demand for skilled labor and conribue o increasing inequaliy. Oher variables used in Eq. (7) are chosen as conrol variables. A paricularly useful conrol variable is he percenage of informaion, echnology, and capial sock in he oal sock of capial since i is an indicaor of echnological developmen, which is expeced o increase inequaliy by raising demand for skilled labor. However, i is challenging o separae he effec of globalizaion and he effec of echnology from each oher, as echnological advances have been shown o help deepen rade and financial links beween counries, while globalizaion has helped spread he use of echnology. A similar concern beween globalizaion and conrol variables arises in he secoral percenage of employmen because globalizaion can move labor away from he agriculural aciviies in he advanced counries. Addiional deail on he esimaes of hese variables are provided in he comparison of he curren sudy s resuls wih hose of IMF in 6. 4 Daa 4.1 The Macro Daa The daa is colleced from he same sources as were used wih IMF WEO (2007). The sample period he IMF chaper uses is This sudy firs replicaes he IMF esimaions and hen uses he exended sample covering 1969 hrough The de faco measure of rade globalizaion ha he IMF sudy uses he sum of impors and expors of goods and services over GDP is from he World Bank's World Developmen Indicaors (WDI) daabase. The de jure measure of rade globalizaion is calculaed as 100 minus he ariff rae, which is he mean of he average unweighed ariff 7 The IMF auhors also run a robusness check by using a logisic ransformaion of he Gini index (making he variable compleely unbounded), ye hey reveal he esimaes obained hrough he use of he log of he variable. Wha I infer from his pracice is ha convering he daa in an aemp o make he variable unbounded did no change heir resuls significanly. 12

13 rae from he WID daabase and of he effecive ariff rae (ariff revenue/impor value) from he OECD saisics daabase. The de faco measure of financial globalizaion is cross-border asses and liabiliies as a percen of GDP from "Exernal Wealh of Naions Mark II" creaed by Lane and Milesi-Ferrei (2006). This percenage includes, for boh asses and liabiliies, FDI, porfolio equiy, deb, financial derivaives, and oal reserves minus gold (asses only). The de jure measure of financial globalizaion is he capial accoun openness index, which is obained from Chinn and Io (2006). To measure counries' financial deph, IMF uses he volume of credi offered o he privae secor by deposi banks and oher financial insiuions as a percen of GDP. The source is he Financial Srucure daabase prepared by Beck, Demirgüç-Kun, and Levine (2000) and revised in March Daa on educaional aainmen of he populaion aged 15+, he populaion s average years of educaion, and he percen of he populaion wih secondary or higher educaion are from Barro-Lee s (2000) daase. The daa is available for bu only in 5-year inervals. For he oher years, I inerpolae he daa linearly, and I use linear exrapolaion for he years afer IMF (2007) uses he capial sock series prepared by Fajnzylber and Lederman (1999), and he ICT (Informaion Communicaion Technology) capial sock series prepared by Jorgensen and Vu (2005) o calculae he share of ICT in he oal capial sock as a measure of echnological developmen. Since his series is no longer available, I use boh he oal and ICT capial sock from he join work of Dale Jorgenson and Khuong Vu of Harvard Universiy and he Toal Economy Growh Accouning Daabase of he Groningen Growh and Developmen Cenre. These series cover only he years afer As a measure of echnological developmen, used in he longer sample esimaes, I use he yearly number of paens graned by he Unied Saes Paen and Trademark Office (USPTO) and he European Paen Office (EPO). This daa is obained from he Innovaion and Developmen Daabase prepared by Laura Saenz and World Bank auhor Daniel Lederman. In replicaing he IMF esimaion, I use daa on he percenage of people employed in agriculure and indusry from he WDI daabase. Since his daa is available only from 1980 onward, I use he daa from OECD saisics daabase, going back o1965, for he exended sample. To calculae labor produciviy in agriculure and indusry (wih respec o oal labor produciviy), I use daa for he percenage of employmen in hese indusries from boh he World Bank and he OECD daabases and use he value-added of hese indusries as a percen of GDP from he World Bank daabase. 4.2 The Micro daa The micro income daa used o consruc wihin-cohor and aggregae inequaliy indices are based on he CPS (Curren Populaion Survey) for he US and on he FES (Family Expendiure Survey) for he UK. The IMF (2007) uses hese sources as well. The CPS 13

14 daa is exraced from IPUMS (Inegraed Public Use Microdaa Series), and he FES daa is downloaded from he ESDS (Economic and Social Daa Service) of he UK governmen, disribued by he UK Daa Archive, Universiy of Essex, Colcheser. Neiher of hese sources supplies panel daa, so he households in hese sources are no he same households ha are followed o calculae evaluaion of wihin-cohor inequaliies. Insead, he synheic wihin-cohor inequaliy daa is consruced from hese cross-secional daa ses. For example, he inequaliy index in 1969 is calculaed for heads of households aged years, and in 1970 he same index is calculaed wih heads of households aged22-26 years. The Gini coefficien is calculaed as shown in Eq. (8): Gini 1 ( n 1 i) yi ( 1 2( i n )) N y i, where he daa is in non-decreasing order: yi y i 1. The micro income daa for he US is he gross moneary income of households, unadjused for household size. The micro income daa for he UK is he disposable income of he households, adjused for household size by following he sandards of HBAI (Households Below Average Income) produced by he UK Deparmen for Work and Pensions. Two ypes of HBAI scales are used o weigh he household members in he adjusmen of household size; I use he one ha FES uses, shown in Table 1 in he Appendix. The daa for he UK excludes Norhern Ireland, so i is acually Grea Briain daa. Froszega e al. (2000) compare he income disribuion of Grea Briain and he UK and show ha he difference is negligible. The FES changed is calendar year in Before 1994, i colleced daa for he previous calendar year, bu in 1994 i began collecing daa a he same ime (March) ha CPS did. The CPS daa, if i covers he period from he 1994 April o 1995 March, is called daa, which I label as 1995 daa, and I did he same for he FES. As a resul, he daa for FES is missing daa for The 1994 daa is consruced by averaging inequaliies in 1993 and 1995 for cohor and aggregae inequaliies. As a final poin, he UK daa covers and he US daa In order o follow he inequaliy of he same cohors while mainaining consisency beween esimaes, he US daa is also used from 1969 unil 2003, where he IMF sudy ends. i (8) 4.3 Analysis of he Daa Evoluion of Aggregae Inequaliy in he US and he UK Figure 8 compares he aggregae inequaliy daa consruced in his sudy wih he World Income Inequaliy Daabase (WIID)-based daa ha is used by he IMF, WEO (2007) sudy. For he US, even hough he levels and rends of boh daa ses are similar, here is a clear difference beween hem in he 1990s. This difference is unexpeced, as 14

15 his daa is consruced from he CPS series, he same source used by WIID. Furhermore, as he original inequaliy daa source (he US Census Bureau, which conducs he CPS) does no adjus for size of he household, I do no eiher. Neverheless, he objecive of deermining wheher using wihin-cohor daa improves esimaes (a leas in erms of sandard errors) compared o he aggregae daa requires only ha he aggregae and wihin-cohor inequaliies be formed from he same source so he resuls can be compared, and his goal is achieved. For he UK, he WIID daa and he daa consruced in his sudy mach nearly perfecly. There is a small difference beween he level of he consruced series and he original series, bu his difference seems o be consan over he years. Therefore, he series are idenical from he poin of view of his sudy. Given he discrepancy in he US daa, he good mach of he UK daa is surprising because he daa producion required downloading he survey daa for each year and hen combining variables o correc for household size. Moreover, he WIID consrucs he UK inequaliy daa using wo daa sources, whereas his sudy uses only one Evoluions of Wihin-Cohor Inequaliies in he US and he UK Figures 9 and 10 show he evoluion of wihin-cohor inequaliies in he US and he UK. This daa is consruced by dividing he populaions ino five-year cohors and using only hose in he populaion who are age The number of households in each cohor and each year are shown in Tables 1 and 2. Figures 9 and 10 reveal a clear paern: wihin-cohor inequaliies increase by age over ime, as prediced. The increase in inequaliy seem o slow down when cohors (or heads of households) move pas age 55, as a around ha age people sar o reire, afer which heir incomes are expeced o level off. 9 The UK daa is no as smooh as he US daa, bu his is undersandable given he limied number of cohor cells used o calculae he inequaliies for he UK, compared o hose used o calculae he US inequaliies. 5 Comparison of Measures of Inequaliy Figure 11 shows he resuls of comparing average growh in wihin-cohor inequaliy wih growh in he aggregae (average wihin-cohor) inequaliy using daa for he US and UK. The graphs use hree measures of inequaliy: (1) he average growh rae of wihin-cohor inequaliies for he heads of households aged 16-65, (2) he growh rae of 8 For he sake of space, he graphs belonging o he las wo cohors, he sample sizes of which are less han 200, are no shown. 9 In he US daa, he wihin-cohor inequaliy also decreases in he iniial periods of he younges cohors, and his paern remains he same across observaions repeaed a differen poins in ime. Alhough i is no wihin he scope of his paper, his observaion may indicae ha, around he ime cohors are expeced o ener he labor force, he inequaliy among hem lessens, possibly because hey were previously living alone and funded by heir families. 15

16 aggregae inequaliy calculaed for he heads of households aged 16-65, and (3) he growh rae of aggregae inequaliy, consruced in accordance wih he WIID daa ha uses he full sample of households. The graphs in Figure 11 poin ou ha aggregae inequaliies calculaed using full sample and he sample of heads of households aged are very close o each oher and ha here is no subsanial difference beween he average growh of wihin-cohor inequaliy and he growh of aggregae inequaliy, especially for he US. This resul seems a firs o conflic wih my claim ha aggregae daa behaves differenly from he daa ha averages wihin-cohor inequaliies. However, hese figures are drawn for 1969 onward, and Figure 7 shows ha he US inequaliy daa keeps increasing for ha period, raher han sabilizing. Therefore, we should expec similar paerns in he measures for he US. On he oher hand, Figure 7 shows ha he inequaliy daa for he UK sabilizes around 1990, and he UK graph in Figure 11 shows ha his is exacly he period where he cohor daa deviaes from he aggregae daa. Anoher explanaion for similariy of he measures is ha he daa for he average growh of wihin-cohor inequaliy in Figure 11 uses cohors consruced from 1969 onwards. Since he younges cohor in 1969 is years old, he younges cohor in 1970 is years old. Therefore, a he ime of he change in aggregae inequaliy in he UK, 1990, he younges people in he younges cohor are 37 years old, and he average household age is around 50 (hose above age 65 are dropped). Hence, he effec of a srucural change in he economies on he change in average growh of cohor inequaliies occurring afer 1970 can easily be underesimaed. To accoun for his problem, he graphs can be correced by using, for example, a rolling window of cohors, which I will address in fuure work. 6 Panel Esimaes 6.1 Summary of he IMF Resuls The IMF WEO (2007) resuls indicae ha echnology in he advanced economies has conribued as much o increasing inequaliy as globalizaion has, whereas oher conrol variables such as he indicaors for financial developmen, educaion level, and changes in he agriculural and indusry employmen shares have miigaing effecs on inequaliy. Decomposing he effec of globalizaion, he IMF auhors find he effec of rade on inequaliy o be negaive ha is, ha i decreases inequaliy. This effec is sronger on he volume of expors han on ariff liberalizaion. The effec of financial liberalizaion is posiive ha is, i increases inequaliy. This effec is sronger wih he ouward FDI han he inward FDI These effecs, which cover , are shown in Table 4.10 and Figures 4.9 and 4.10 of he IMF publicaion. 16

17 6.2 Re-Esimaing he IMF Esimaes The IMF summary model, in Table 4.10 of IMF WEO (2007), is re-esimaed for he US and he UK, firs separaely, hen ogeher, hrough panel regressions by using aggregae inequaliy indices. This re-esimaion obains a subse of he resuls he IMF obains and uses he same inequaliy daa source he IMF uses, WIID. The resuls are shown in he firs columns of Tables 3, 4, and 5. Even hough I propose he growh rae of wihin-cohor inequaliy as he righ measure o use, I do no immediaely change he independen variable from he level of he WIID-based aggregae index o he growh rae of wihin-cohor inequaliy bu do a sepwise procedure and make a one change a a ime. The second columns of Tables 3, 4, and 5 reveal he resuls of esimaions using he growh rae of he WIID Gini index daa, he hird columns use he growh rae of aggregae Gini index daa I consruced, and he fourh columns use he growh rae of wihin-cohor inequaliy indexes I consruced. The esimaions on he firs hree columns of Tables 3, 4, and 5 use he sample size of 22 years, as he IMF did. However, while he IMF sudy uses en explanaory variables, his sudy canno do he same because I use only wo counries, and employing en explanaory variables a a ime would leave he regressions wih very few degrees of freedom. Insead, I proceed in an alernaive way: The IMF uses wo conrol variables o measure he effec of channels oher han rade ha may affec inequaliy, so one ou of wo conrol variables is eliminaed. In deciding which of he wo o eliminae, I chose he less significan variable on he IMF able. In he end, I eliminaed he percenage of populaion wih a leas secondary educaion in favor of he average year of schooling, and he percenage of agriculural employmen in favor of he percenage of indusry employmen (even hough hese are no complemens). Finally, recognizing ha he de jure measure of financial globalizaion he IMF uses he capial accoun openness index says consan for he US and he UK during he sample period, I exclude his variable as well. In he end, he number of explanaory variables is reduced o seven, for which I use he same daa conversion as he IMF used: he naural logarihm of all he dependen and independen variables, excep for he 100 minus ariff rae and he percenage of he populaion wih a leas secondary educaion. Unforunaely, neiher he IMF sample period ( ) nor he ICT daa used by he IMF WEO (2007) o measure echnological developmen exends before he 1980s, so I canno idenify he changes in daa ha occurred in he 1970s in he UK daa (Figure 7). For his reason, I exend he daa by replacing he ICT daa wih he number of paens graned. However, looking a he graphs and focusing on only some pars of he daa would consiue daa mining, so for he curren purposes, I sick wih he IMF sample and use he exended sample in a robusness check. 17

18 In Tables 3, 4, and 5, he panel esimaes are compared wih ime series regression esimaes. Even hough i is sandard o use year dummies in he panel esimaions, hey do no appear in he ime series regressions, so hey are employed in he robusness checks bu no in he esimaes of hese ables. 6.3 Resuls Table 3, which shows he resuls for he US, indicaes ha, when he level from he WI- ID-based Gini index is used, boh rade volume and financial liberalizaion increase inequaliy, bu his resul is significan only a he en percen level. However, hese esimaes do no have be idenical o he IMF esimaes, as only wo counries are used. The quesion is wheher he proposed inequaliy index will improve hese indices. The sign of he coefficien of rade changes when he growh raes of he WIID Gini or he Gini I consruced is used, bu he resuls are insignifican. These wo se of resuls mach reasonably well. Finally, when he growh rae of wihin-cohor inequaliies is used, boh he rade and financial globalizaions lead significanly o an increase in income inequaliy. The ICT capial share is also srongly associaed wih increasing inequaliy, while financial developmen is negaively associaed wih increasing inequaliy. Oher esimaes have he expeced signs, bu hey are no significan. Resuls for he UK are shown in Table 4. The firs column, using he level of aggregae inequaliy, finds a miigaing effec of rade and a worsening effec of financial liberalizaion on inequaliy, jus as he IMF resuls do. However, using he growh rae of hese variables in columns 2 and 3 and especially using he growh rae of wihincohor indices in he las column causes he signs of coefficiens of hese variables and he signs of mos of he conrol variables o change. Given he paern of he UK daa, his change in he resuls is no surprising. In 1973 he rade share increases suddenly, and he fac ha inequaliy coninues o increase from ha ime on occurs probably because of he cohor srucure of he populaion. Finally, Table 5 shows he esimaes using boh counries a once. As expeced, he resuls are close o being he average of Tables 3 and 4. The esimaes in he las column of Table 5 are used as baseline esimae for he robusness checks in he nex secion. The coefficien of rade share in GDP is posiive and significan; ha is, i has an aggravaing effec on GDP. The effec of he oher rade liberalizaion indicaor, ariff rae, is he same, alhough i is insignifican. The coefficien of he raio of inward FDI sock o GDP is negaive and insignifican and much smaller han ha of he rade share. The coefficien of he raio of informaion, communicaion, and echnology capial o oal capial sock is posiive and highly significan, so he skill-biased echnological change explanaion for he rise in earnings inequaliy finds srong suppor. The coefficien of financial developmen, measured by he raio of credi offered o he privae secor as a percen of 18

19 GDP, is negaive bu is only significan a he 10% level. The coefficien of he average years of educaion is significan and negaive; his resul is expeced since educaion is par of skill, and an increase in he level of educaion is an indicaor of he increase he number of skilled labor, which reduces inequaliy. The final variable is he percenage of employmen in he indusry secor, which has posiive and insignifican sign, he only unexpeced sign found. 6.4 Robusness Check This secion presens hree robusness checks of he resuls ha are in he las column of Table 5. Henceforh, he resuls in he las column of Table 5 will be called he base esimae. Cohor and counry dummies are used in his esimaion. The firs robusness check (Table 6) changes he conrol variables. The firs column of Table 6 shows he base esimae. The second and hird columns change he indicaor of financial liberalizaion from he Inward FDI sock as a percen of GDP o Inward FDI flow and Cross-Border Asses and Liabiliies as percens of GDP, respecively. The coefficiens of hese variables say nearly he same across regressions ha is, negaive and insignifican. Changing he liberalizaion indicaors does no cause a noiceable change in he coefficiens of oher variables. Since using wihin-cohor daa insead of he aggregae daa provides more observaions and, hence, more degrees of freedom in he esimaes, i is possible o use he oher conrol variables IMF uses ha are no included in he base esimae. The resuls are on he fourh column. Mos of he resuls for he variables ha are also used by he base esimae are similar o he base esimae. For he newly added variables, resuls indicae ha he percenage of he populaion wih a leas a secondary educaion has a negaive bu significan effec on inequaliy. However, his resul may have been affeced from he high correlaion of his variable wih he oher measure of educaion, average years of schooling. The oher newly added variable is he percenage of agriculural employmen, which has a miigaing effec on inequaliy. This resul is unexpeced, as is he effec of he percenage of indusry employmen on inequaliy. Table 7 shows he resuls of he second robusness check changing he sample period and Table 8 presens he resuls of he hird robusness check running he same regression wih he base esimae wih he addiion of year dummies and hen counrycohor dummies. Adding counry-cohor dummies causes almos no change in eiher he coefficien esimaes or he sandard errors. However, addiion of year dummies improves he resuls significanly in erms of sandard errors. Oher han he coefficien of he ariff rae, all coefficiens preserve he signs. These resuls indicae ha rade liberalizaion causes inequaliy o increase, and financial liberalizaion causes i o decline. Overall, he robusness checks provide considerable suppor for he conclusions drawn 19

20 from he base esimae: Trade has led o increase in he income inequaliies in he US and he UK, and given he rade effec, he effec of financial liberalizaion is negaive and insignifican. When rade share is no included in GDP bu is only an indicaor of financial liberalizaion, is coefficien is insignifican bu posiive. This change may occur because of he high level of correlaion beween he variables. 7 Concluding Remarks This sudy began by discussing which measure of inequaliy should be used in regression esimaes, and his par concludes ha he righ measure is he average growh rae of wihin-cohor inequaliies. The sudy hen revisied he relaionship beween inequaliy and rade using US and UK daa and found ha, conrary o he negaive and insignifican relaionship beween he variables obained by regressions found hrough he use of aggregae inequaliy daa, he relaionship beween he variables is posiive and significan when he proposed measure of inequaliy is used. This resul is consisen wih rade heory and a subsanial par of he rade lieraure. The changes in wihin-cohor inequaliies have been explained by changes in he disribuion of individuals' characerisics, such as experience and educaion, bu o he bes of my knowledge, his is he firs sudy ha relaes hem o variables like rade. The cohor esimaes reveal ha he average number of years of educaion is inversely relaed o inequaliy, and echnological advances and financial developmen increase inequaliy. The effec of financial liberalizaion is no robus o a change in he variable chosen o represen i. Inward FDI as a percen of GDP seems o increase inequaliy, bu crossborder asses and liabiliies as a percen of GDP has he opposie effec. I is difficul o differeniae he effecs of rade and financial liberalizaion hrough he regression esimaes, as rade and financial liberalizaion are highly correlaed. However, he resuls for he increasing effec of rade on inequaliy remain valid in he se of robusness ess. Generally speaking, even hough rade is found o have a worsening effec on inequaliy, many advances, such as echnology diffusion and he inernaional movemens of facors of producion, occur a he same ime as an increase in rade globalizaion, which may have caused he change in inequaliy. Oher domesically oriened explanaions for he changes in inequaliy may also play a role following he inegraion of a counry wih he res of he world. Absracing as much as possible from hese discussions, my aim in his sudy is o bring o aenion o he coincidence of changes in inequaliy wih large changes in rade and financial indicaors and o he inabiliy of he relaionship beween rade and inequaliy o be discerned based on simple regression esimaes. 20