The own-skill-group effect of emigration on source country real wages: An analysis of Jamaican emigration Thesis submitted in partial fulfilment of the requirements for the Degree of Master of Science in Economics for Development at the University of Oxford Dan Brown Supervisor: Professor Paul Collier Date: 01/06/12 Candidate Number: 206416 Word Count: 9,678 (Excluding title page, abstract and references: 31 pages, with 309 words on an average ‘all-text’ page (p.31) + 99 words on p.33) Abstract This extended essay contributes to the nascent literature investigating the ownskill-group effect of emigration on real wages in source developing countries. After presenting a new, simple neo-classical general equilibrium theoretical model to distinguish own-skill-group from cross-skill-group real wage effects and their likely sign, it augments the existing empirical analysis by using a more precise skill-group classification and more thoroughly evaluating the robustness of estimates; considering the previously unstudied case of Jamaican emigration. Fixed effects two-stage-least-squares estimates suggest an emigration-induced 10% reduction in the Jamaican labour force in a skill-group increases real wages of remaining workers in that skill-group by 4.34%. 1. Introduction Barriers to international migration perhaps constitute the largest remaining distortion to an increasingly integrated world economy. As such, the associated global efficiency losses are huge; Clemens claims that ‘when it comes to policies that restrict emigration, there appear to be trillion dollar bills on the sidewalk’ (p.1 Clemens (2011)). Clemens (2011) presents a simple diagrammatic analysis of the effects of migration of workers from a low-wage developing to high-wage developed country in a neo-classical two-factor (‘land’ and homogeneous ‘labour’) model. Assuming diminishing returns to labour: 2 Figure 1: Clemens’ diagrammatic analysis: A relaxation of international migration barriers such that emigration results in a reduction of the developing country labour force from L1 to L2 brings a gain of area c+b to the migrants themselves (receiving a real wage MPLH2 rather than MPLL1), a loss of area e to developed country workers from the lower real wage MPLH2 (from MPLH1), a gain of area e+d to owners of the fixed factor land in the developed country, losses of a+b to developing country landowners, and gains of area a to developing country workers through the higher real wage MPLL2 (from MPLL1). We therefore observe net global efficiency gains of area (c+d), alongside distributional effects between owners of different factors of production in both the developed and developing country. The size of the net global efficiency gains is increasingly well documented. For example (with respect to the size of area c), Clemens, Montenegro and Pritchett (2010) find that a 35-year-old urban male Jamaican worker with 9 years of Jamaican education could earn between 329 and 363% of his wage by moving to the US. The literature regarding the distributional effects of immigration for recipient developed countries is even vaster. 3 By contrast, the part of Clemens’ diagram that remains significantly underresearched is the distributional effect of emigration across owners of factors of production in the source developing country. It is to this end that this extended essay is focussed. Specifically, it attempts to understand the effect of emigration on the real wages of remaining workers of the same ‘skill-group’ (defined in terms of a individual’s education, experience and occupation): the ‘own-skillgroup’ effect of emigration. The sparse existing literature has considered source country real wage effects of emigration in Mexico, Moldova, Honduras, Puerto Rico and Lithuania. As yet, the own-skill group effect in Jamaica remains unstudied. Substantial emigration from Jamaica, particularly to the US, UK and Canada, has resulted in a stock of Jamaicans representing 35.77%1 of the remaining Jamaican population living outside the country in 2000 (World Bank). This is amongst the highest emigrant stocks in the world, and is considerably higher than in Mexico, the focus of the existing literature (9.56%1). In this extended essay, I augment the empirical approach taken in previous analyses and extend it to the case of Jamaican emigration. Section 2 summarises the existing literature on the real wage effects of emigration in source countries. Section 3 presents my new theoretical model of those effects, whilst section 4 presents my empirical analysis. The essay concludes in section 5 by interpreting the economic magnitude of the estimated own-skill-group real wage effects for poverty reduction among the wage-earning poor who have little ownership of fixed factors of production. 2. Existing literature The entire existing empirical literature is limited to a handful of papers. The three earliest papers provide evidence that is at best consistent with the existence of a positive effect of emigration on remaining workers’ wages. Lucas (1987) estimates a system of equations exploiting variation in numbers of 1Dividing total emigrants, 2000, in the World Bank bilateral migration database by total population from the World Development Indicators 4 emigrants over time from Malawi and Mozambique to work in South Africa’s mines, over the period 1946-78. He identifies a positive and significant effect of emigration on the real wages of plantation workers in both source countries. This allegedly played a role in the British decision to introduce quota controls on the movement of migrants from colonial Nyasaland (modern day Malawi) to South Africa’s mines, to protect white European plantation owners from rising labour costs. Boyer et al. (1993) present descriptive statistics specifically consistent with a positive own-skill-group effect of emigration. They find that two types of unskilled real wages increased dramatically in Ireland in the absence of industrialisation over the period 1860-1913 at the same time as mass emigration of unskilled workers to US and Britain. Real agricultural labourer wages doubled, bricklayers’ wages increased by a factor of 2.4, whilst the rural population fell by 2.6 million. Finally, Hanson (2005) uses a difference in differences estimator to demonstrate higher real wage growth 1990-2000 in those Mexican states with ‘high’ rates of emigration in 1950. Mishra (2007) initiated a more rigorous empirical approach to identification of the own-skill-group effect, borrowing the ‘national approach’ first presented in the impact of immigration literature in Borjas (2003). The national approach uses variation in the emigrant share2 and source country wages across ‘skill groups’, defined in terms of education and labour market experience, to identify the effect. Every paper in the literature has identified a positive and significant own-skill-group effect; typically of the order of magnitude of a several percentage point increase in source country wages for a 10% increase in the emigrant share. In her benchmark fixed effects regressions, Mishra finds that in the case of Mexican emigration to the US 1970-2000, a 10% increase in the emigrant share raises real wages of workers in the same skill group by 4.4%. To deal with the possibility of reverse causality (the fact that low domestic real wages may drive emigration, discussed in section 4b)iv)), she uses the lagged average emigrant share for workers of the same education group across all experience groups as an instrument for the emigrant share. The validity of this, and other, instruments 2Emigrant share taken as total number of emigrants divided by total labour force in the source country. 5 employed in the literature is discussed in section 4b)iv). In her IV regressions she unsurprisingly finds the positive and significant effect increases. Over the period 1960-2000, Aydemir and Borjas (2006) find a 10% increase in the emigrant share is associated with a 5.6% increase in remaining workers’ real wages in the same skill group in Mexico, although they do not attempt to uncover a causal effect. Borjas (2007) finds that same shock to the emigrant share relates to a 2.1% increase in own-skill-group real wages for Puerto Rican emigration to the US 1970-2000. He attempts to identify a causal effect using variation in US average wages across skill groups as an instrument for the emigrant share, and finds an economically larger effect. Gagnon (2011) also uses the US average wages instrument in the case of Honduran emigration to the US 2001-7, and finds a 12.8% increase in own-skill-group real wages for the 10% increase in emigrant share. Bouton (2011) uses household survey data from Moldova for a single cross-section 2006 to find a 4% increase in own-skill-group real wages associated with this emigrant share shock in his male sample, but this is only a correlation. Finally, (again following Borjas (2003)) some papers have simulated the total effects of emigration (incorporating both own-skill-group and cross-skill-group effects, as explained in section 3) based on a structural model of the source country economy calibrated with estimated elasticities of substitution between workers of different types. Elsner (2011) claims that Lithuanian emigration 2002-6 increased the real wages of workers with less than 10 years experience by 4.9-7%, and reduced real wages of workers with more than 30 years experience by 1%. The Aydemir and Borjas paper does similarly, but with conclusions vulnerable to parameter changes in the model. 3. Theoretical model Bouton (2011) is the only of the aforementioned papers to provide an explicit theoretical model of the effects of emigration on real wages in source countries. However, he treats labour as a homogeneous factor input, and as such cannot 6 distinguish between own-skill-group and cross-skill-group effects. As described above, a structural model is presented in Elsner (2011) and Aydemir and Borjas (2006), but given its empirical purpose, to produce simulations of the overall effects of emigration, its complexity detracts from a clearer understanding of the intuition underlying the real wage effects of emigration. I therefore here present my new, simple neo-classical general equilibrium model of the source country economy, the essence of which is similar to that structural model. i) The model: Suppose that output, Y, is produced by a combination of capital, K (alternatively K could represent land), and ‘aggregate’ labour, L, according to a constant returns to scale, Cobb-Douglas production function. Hicks-neutral technology is denoted by A, and α represents labour’s share of the value of output, where 0 < α < 1: ππ‘ = π΄π‘ πΎπ‘1−πΌ πΏπΌπ‘ (1) Aggregate labour is in turn a composite of the two types of worker that exist in the economy, skilled workers, S, and unskilled workers, U, according to a constant elasticity of substitution production function; where it is assumed that the γi > 0 (i = 1, 2). The parameter ρ = (σ – 1)/σ, where σ denotes the elasticity of substitution between skilled and unskilled workers in the production of aggregate labour. I restrict attention to cases where the CES production function is convex, thus 1 ≥ ρ ≥ -∞ (where ρ = 1 gives the case where skilled and unskilled workers are perfect substitutes, and ρ = -∞ where they are perfect complements): π π 1 πΏπ‘ = (πΎ1 ππ‘ + πΎ2 ππ‘ )π (2) Substituting (2) into (1) and differentiating with respect to S, I derive an expression for the marginal product of skilled labour, which gives the real wage of skilled workers under perfect competition: 7 πππ‘ πππ‘ π−1 = πΌ πΎ1 ππ‘ π π πΌ π΄π‘ πΎπ‘1−πΌ (πΎ1 ππ‘ + πΎ2 ππ‘ )π −1 (3) ii) The own-skill-group effect: Differentiating (3) by the stock of skilled workers, S, tells us how an emigration induced decrease in the stock of skilled workers affects the real wages of remaining skilled workers: the own-skill-group effect of emigration for skilled workers (the expression for the case of unskilled workers is symmetrical). π2 ππ‘ = πππ‘2 π−1 [πΌπΎ1 ππ‘ π΄π‘ πΎπ‘1−πΌ ] [(πΌ π−2 [πΌπΎ1 (π − 1)ππ‘ − π−1 π π)(πΎ1 ππ‘ )(πΎ1 ππ‘ π πΌ + π π΄π‘ πΎπ‘1−πΌ (πΎ1 ππ‘ + πΎ2 ππ‘ ) π −2 πΎ2 ππ‘ )π ] + πΌ −1 π (4) ] Factorising this expression and substituting in Lt for notational clarity: π2 ππ‘ πππ‘2 π−2 πΌ−π πΏπ‘ ][πΎ1 (πΌ = [πΌπΎ1 π΄π‘ πΎπ‘1−πΌ ππ‘ π −π − π)ππ‘ πΏπ‘ + (π − 1)] (5) Given that S, L, K and A are positive quantities by definition, and γ1 > 0, 0 < α < 1, the term in the first set of brackets is positive. The sign of the own-skill-group effect therefore depends on the sign of the expression in the second set of brackets: π2 ππ‘ πππ‘2 π < 0 if: [πΎ1 (πΌ − π)(πΏπ‘ )π + (π − 1)] < 0 π‘ π i.e. if: πΎ1 (πΌ − π)(πΏπ‘ )π < (1 − π) π‘ Case 1: α ≥ ρ: Firstly, suppose α ≥ ρ. In this case, (α – ρ) ≥ 0, thus: 8 π2 ππ‘ πππ‘2 < 0, ππ: π πΎ1 (πΏπ‘ )π < π‘ (1−π) (πΌ−π) (6) Since 0 < α < 1, the numerator of the right hand side of (6) exceeds the denominator, thus the right hand side exceeds 1. Raising (2) to the power ρ: π π π πΏπ‘ = (πΎ1 ππ‘ + πΎ2 ππ‘ ) π Dividing both sides by ππ‘ : πΏπ‘ ππ‘ ( )π = πΎ1 + πΎ2 ( )π ππ‘ ππ‘ (7) Thus, given γ2, Ut, St > 0: πΏπ‘ ( )π > πΎ1 ππ‘ Thus: ππ‘ (1 − π) πΎ1 ( )π < 1 < πΏπ‘ (πΌ − π) Therefore, inequality (6) holds. Case 2: α < ρ: Instead suppose that α < ρ, such that (α – ρ) < 0. In this case: π2 ππ‘ < 0, ππ: πππ‘2 ππ‘ (1 − π) πΎ1 ( )π > πΏπ‘ (πΌ − π) (8) 9 Given that 1 ≥ ρ, the numerator is zero or positive, and the denominator negative, and thus the right hand side of (8) is zero or negative. Thus given γ1, Ut, St > 0, the inequality holds. I have arrived at the unambiguous prediction that π2 ππ‘ πππ‘2 < 0: ceteris paribus emigration raises the real wages of remaining workers of the same skill group. iii) The cross-skill-group effect: Differentiating (3) by the stock of unskilled workers, U, tells us how emigration of unskilled workers affects the real wages of remaining skilled workers; the cross-skill-group effect of unskilled emigration on skilled workers: π2 ππ‘ π−1 πππ‘ πππ‘ = πΌπΎ1 πΎ2 (πΌ − π)ππ‘ π−1 ππ‘ π π πΌ π΄π‘ πΎπ‘1−πΌ (πΎ1 ππ‘ + πΎ2 ππ‘ )π −2 (9) The sign of the cross-skill-group effect therefore depends on the sign of (α – ρ). If π2 π α > ρ, ππ πππ‘ > 0, and thus emigration reduces the real wages of remaining π‘ π‘ workers of the other skill group. By contrast if α < ρ, emigration increases the real wages of remaining workers of the other skill group. iv) Intuition: Diminishing marginal returns to factor inputs drives the unambiguous own-skillgroup effect. Quite simply an emigration-induced reduction in the supply of workers identical to you raises your marginal product, and so in perfectly competitive markets your real wage. The cross-skill-group effect is ambiguous, and crucially depends on the size of capital’s share of the value of output (1-α) relative to the degree of complementarity between skilled and unskilled workers as captured by ρ. Emigration of a worker more complementary in production to you (lower ρ) intuitively has a more harmful effect on your marginal product, and so increases the likelihood that α > ρ, i.e. your real wage reduces. Your real 10 wage is also more likely to reduce as capital’s share of the value of output is lower, thus α higher. Intuitively, the lower is the relative importance of capital in production, the less a reduction in aggregate labour (from emigration of workers of the other skill group) will translate to an increase in the marginal product of aggregate labour, and so in turn through to the marginal product (thus real wages) of workers of both types. The model provides the essential function of explicitly noting the two separate effects of emigration. The total effect of emigration for a skill group depends on the sum of the own-skill-group and all cross-skill-group effects on that skill group. A failure to distinguish effects has led some authors to draw misguided conclusions about the overall impact of emigration having only estimated an own-skill-group effect. For example, Mishra (2007) uses her estimate of the ownskill-group effect combined with relative changes in earnings shares of different skill-groups to ‘aggregate’ wage impacts across all skill-groups in Mexico, claiming that emigration 1970-2000 raised the average Mexican worker’s wage by 8%. This completely ignores the existence of cross-skill-group effects. Not only do I refrain from making these statements, given that I only attempt to estimate the own-skill-group effect in the next section, but the theoretical model demonstrates the importance of controlling for cross-skill-group effects in my regression analysis if I want to understand the true own-skill-group effect. v) Extensions to the model: Departure from full employment: Katseli et al. (2006) note that the effects of emigration depend on source country labour market conditions. So far the model has assumed full employment. However, in the presence of unemployment the effects of emigration may operate through changes in the unemployment rate rather than real wages. Emigration of unemployed workers from a skill group, or emigration of employed workers who are subsequently replaced by identical unemployed workers, may have no own or cross-skill-group effects on real wages since relative supplies of employed factor inputs remain unchanged. This 11 may be important in the Jamaican context; Kim (2007) highlights double-digit unemployment for much of the 1990s and 2000s. Short run versus long run effects: The model looks at short run effects during which time the supply of all other factor inputs as well as production methods remain fixed. In the long run the economy may adjust to an emigration-induced change in factor endowments through three channels other than changes in factor prices. Dustmann and Glitz (2011) outline two of these in a HeckscherOhlin-Samuelson model. Firstly, the source country may begin to produce a higher quantity of the good whose production is relatively intensive in those factors that become increasingly abundant due to emigration (the Rybczynski effect). Secondly, within the existing structure of production firms across all sectors may begin to employ production techniques that make greater use of those now increasingly abundant factors. Finally, lower returns to capital may discourage capital investment. All three effects will supress the own-skill group real wage increase from emigration in the long run. Less capital implies the relative scarcity of those types of labour in which there has been emigration, which drove the real wage increase, is reduced. The Rycbzynski effect and the change in production techniques imply a reduction in demand for the emigration-induced scarcer workers, so again supressing their real wage increases. The long-run cross-skill group effects can, however, differ from the short-run in either direction. For example, suppose workers in skill-group A suffer an adverse cross-skill group effect in the presence of strong complementarities between them and the emigrating workers in skill-group B. That adverse effect may be worsened by the decrease in formation of (complementary) capital. Yet the Rybczynski effect and shift in production techniques may increase demand for workers from skillgroup A, so reducing the adverse effect on their real wages. 12 4. Empirical Analysis a) Data and methodology i) Data sources: Data sources and procedures applied to the raw data are summarised in Table 2 (appendix), but attention is drawn to the following. I use the IPUMS International dataset, in which internationally comparable census data is available for selected countries. Jamaican censuses 1982, 1991, and 2001 are used to obtain information on annual wage income (from all types of employment) in the source country; whilst US censuses 1980, 1990, and 2000 are used to track the number of Jamaican emigrants to the US. Since other Jamaican migrant receiving countries do not publish information about the exact birthplace of residents, only Jamaican emigrants to the US can be counted. The potential implications of this for sample selection bias are discussed in section 4ci). ii) The pattern of Jamaican emigrant stocks 1980-2000: Before turning to the IPUMS data, it is useful to demonstrate the broader pattern of Jamaican emigration 1980-2000, by education and geographically. The Trade Team at the World Bank produces a panel data set documenting changes in emigrant numbers to six OECD countries (UK, US, Canada, Australia, France and Germany) 1975-2000, distinguishing by education levels. Over the period 19802000 there has been a dramatic increase in high-educated (post-secondary education completed) Jamaican emigration. The high-educated share of total Jamaican emigrants to these six OECD countries 1980-2000 increased by 20.71%, nearly doubling its 1980 level of 22.08%. By contrast, the low-educated share (less than upper-secondary completed) decreased by 29.42%, more than halving its 1980 level. According to the World Bank Bilateral Migration Database, of the stock of Jamaican emigrants in 2000, 62.17% were located in the US, 16.16% in the UK, and 12.83% in Canada. 13 iii) Methodology: Following Mishra (2007) as described in section 2, I divide Jamaican emigrants and the ‘remaining’ labour supply (those who stay in Jamaica) into ‘skill groups’. Each skill group defines a type of workers who directly compete. The existing literature has defined skill groups in terms of education and a, crude, measure of labour market experience taken as an individual’s age minus the year they are expected to have joined the labour force (which I take as one year after their completed level of education, with a minimum age of 15). In all regressions four education classifications are used: ‘less than primary education completed’, ‘primary completed’, ‘secondary completed’, and ‘university completed’. For the four respectively, I take workers as entering the labour force aged 15, 15, 19 and 22 for the purpose of the ‘experience’ calculation. I calculate ‘experience’ as ‘age – labour force entry age’, and then in turn use either four 10-year experience group classifications (0-9, 10-19, 20-29, 30+), or eight 5-year classifications (0-4, 5-9, 10-14, 15-19, 20-24, 25-29, 30-34, 35+). Given women are on average more transient in the labour force, defining experience becomes increasingly misleading, and so throughout attention is focussed on males only. I augment the existing literature by, for the first time, using an individual’s ‘primary occupation’ to classify workers at a level of detail beyond education and experience. The ‘OCCISCO’ code in IPUMS defines workers into 10 primary occupations, including, for example, ‘professionals’, and ‘plant and machine operators and assemblers’. In my preferred regressions I therefore divide workers into ‘education-experience-occupation’ skill groups. For each skill group I calculate the emigrant share, defined as the total number of emigrants divided by the remaining labour force (both employed and unemployed individuals). I deflate the nominal wages of employed workers according to the Jamaican CPI from the IMF International Financial Statistics to give real wages, before taking the natural logarithm and calculating the mean across all workers in each skill group. I have a panel of data at the skill-group 14 level over three time periods (approximately 1980, 1990 and 2000 depending on the exact census year). My basic regression equation is a two-way error components model of the following form: πππΊππππ‘ = π0 + π1 πΈππΊπ ππππ‘ + πΏπ πΈππΊπ π πππ‘ + π2 π2 + π3 π3 + π4 πΈπ·π π2 + π5 πΈπ·π π3 + π6 πΈππ π2 + π7 πΈππ π3 + π8 ππΆπΆπ π2 + π9 ππΆπΆπ π3 + πΈπ·π πΈππ ππΆπΆπ + π’ππππ‘ That is to say I regress the mean log real wage in skill-group ijk (education group i, experience group j, occupation group k) at time t (MWGijkt) on the emigrant share in that skill group (EMGRijkt), controlling for the emigrant share in all skill groups of the same experience and occupation level, but different education level, at time t (EMGRsjkt for all s ≠ i) (the cross-skill-effects I control for). Whilst I could have controlled for emigrant shares in skill groups with different experience or occupation but the same level of education, perhaps the strongest complementarities in production occur between workers of different education levels, following Mishra (2007). Mishra (2007) is the only other paper to control for cross-skill-group effects, and even here in only one of her regressions as a ‘robustness check’. I control for time fixed effects through dummy variables for periods 2 and 3 (T2 and T3), to capture variables that influence mean real wages and vary over time but not across skill groups. I control for interactions between education group and the time dummy variables (EDiTt), as well as between experience group and occupation group and the time dummy variables, respectively (EXjTt) and (OCCkTt). These capture all variables causing changes in mean real wages over time by education, experience, or occupation groups. Finally, given there are many plausible unobserved time invariant determinants of mean real wages across skill-groups I wish to control for, I estimate my regressions using fixed effects estimation: thus controlling for EDiEXjOCCk. A Generalised Hausman test, augmenting random effects estimates of specification 15 [2] below with time demeaned variables (regression not reported), gave statistical justification for the use of fixed effects: the null hypothesis that the time demeaned variables had no significant explanatory power was rejected at all reasonable significance levels. The coefficient of interest is θ1, the own-skill-group effect of emigration on mean log real wages of remaining Jamaican workers. It tells us the percentage change in mean log real wages for a 1% decrease in the Jamaican labour supply due to emigration. It can therefore be interpreted as a wage elasticity. Since real wages are simultaneously determined by labour supply and demand, I must control for all influences from labour demand to estimate the real wage effect of the labour supply shock. The large set of fixed effects and interaction terms imply only components of labour demand that vary over time as well as across education-experience-occupation groups (or two of the three) could be contained in the uijkt, and so bias the estimated coefficient of interest if correlated with the labour supply shock. The dependent variable is an average, calculated from a different number of observations in each skill group depending on the number of remaining employed Jamaican workers in that skill group. Each observation for the dependent variable therefore contains information of varying reliability. To take this into account, I weight all regressions using the ‘aweights’ function in Stata. The more individual observations used to calculate the average, the lower the variance of that averaged observation, and so the higher its weight in the regressions. Whilst failing to weight my regressions would not lead to biased estimates, the estimates would, in failing to take into account the information contained in the varying reliability of observations, be inefficient. 16 b) Results i) Descriptive statistics: Before launching into the regression analysis, I present a scatter diagram demonstrating the correlation between mean log real wages and emigrant shares across skill groups across all three time periods. I use the educationexperience classification of the existing literature for visual clarity (given there are far fewer observations by this two-way classification). The size of the points on the scatter graph is given by the number of Jamaican remaining workers used to calculate mean wages; thus larger points represent greater reliability of the observation. Figure 2 plots all observations, before figure 3 ‘zooms in’ on the cluster of observations with an emigrant share less than 0.4: Figure 2: Scatter diagram – Mean log real Jamaican wage against emigrant share by education-experience classification: All observations 17 Figure 3: Scatter diagram – Mean log real Jamaican wage against emigrant share by education-experience classification: Emigrant shares < 0.4 At first glance, mean log real wages appear positively correlated with the emigrant share. Were low mean wages driving Jamaican emigration, I might expect a negative correlation. The descriptive statistics seem sufficiently consistent with a positive own-skill-group effect of emigration to motivate a more rigorous empirical analysis. ii) Results table: Using the specification in [1] I test for serial correlation (using the first difference approach discussed by Wooldridge (2002)), and find that the null hypothesis of no serial correlation is strongly rejected, p-value for the F statistic 0.0037. As suggested in Bertrand, Duflo and Mullainathan (2004), in all the following regressions I cluster standard errors on the skill group (i.e. allow for serial correlation over time within skill groups). 18 Table 1: Empirical results Variable name [1] Fixed effects. Ed-ex skill group. Time invariant weights [2] Fixed effects. Ed-ex-occ skill group. Time invariant weights Constant 11.215*** (0.086) 0.426*** (0.061) 0.245 (0.371) 11.523*** (0.023) 0.002 (0.014) -0.116** (0.058) 0.514*** (0.188) -0.126 (0.118) 0.173 (0.133) 0.434*** (0.126) -0.145** (0.070) 0.281*** (0.063) -0.027 (0.050) 0.668*** (0.181) 1.384*** (0.156) -0.274*** (0.060) -0.377*** (0.053) -0.059** -0.022 (0.014) -0.002 (0.001) 0.541*** (0.152) 1.121*** (0.205) -0.207*** (0.032) -0.233*** (0.045) -6.87x10-4 0.069** (0.033) -1.10x10-4 (0.003) 1.103*** (0.354) 0.423 (0.301) -0.386*** (0.100) -0.109* (0.065) -0.038 -0.138*** (0.048) -0.002 (0.005) -0.750** (0.369) 1.088** (0.428) 0.256** (0.099) -0.243** (0.098) 0.050 0.022 (0.023) -7.28x10-4 (0.002) 0.799*** (0.248) 0.462* (0.276) -0.311*** (0.065) -0.098* (0.058) -0.035 EMGRijkt Cross-skill-group control 1 Cross-skill-group control 2 Cross-skill-group control 3 T2 T3 EDiT2 EDiT3 EXjT2 [3] Fixed effects-2SLS. Ed-ex-occ skill group. Time invariant weights [4] First stage of fixed effects 2SLS regression [5]. [5] Fixed effects2SLS. Ed-ex-occ skill group. Time variant weights [6] Fixed effects. Edex skill group. Time invariant weights. Short versus long run. [7] Fixed effects2SLS. Ed-ex-occ skill group. Time variant weights. Emigrant share adjustment 1. [8] Fixed effects2SLS. Ed-ex-occ skill group. Time variant weights. Emigrant share adjustment 2. [9] Fixed effects2SLS. Ed-ex-occ skill group, 8 experience groups. Time variant weights. 0.256 (0.389) 0.426*** (0.123) -0.142** (0.069) 0.456*** (0.139) -0.152** (0.074) 0.461*** (0.117) -0.064 (0.061) 0.268*** (0.063) -0.036 (0.049) 0.669*** (0.178) 1.443*** (0.175) -0.257*** (0.057) -0.393*** (0.057) -0.060** 0.022 (0.024) -7.27x10-4 (0.002) 0.798*** (0.248) 0.468* (0.273) -0.308*** (0.065) -0.099* (0.058) -0.034 0.034 (0.026) -9.51x10-4 (0.003) 0.824*** (0.262) 0.370 (0.297) -0.314*** (0.068) -0.080 (0.062) -0.039 0.038** (0.018) -3.11x10-5 (0.002) 0.955*** (0.255) 0.405 (0.267) -0.351*** (0.066) -0.107* (0.062) -0.035 11.241*** (0.083) - 19 (0.023) -0.004 (0.022) EXjT3 OCCkT2 OCCkT3 (0.031) -0.056* (0.034) -0.005 (0.010) -9.93x10-4 (0.016) BRAMWGijkt (0.061) 0.027 (0.040) -0.038 (0.024) 0.049** (0.023) -7.86x10-7 (3.38x10-6) (0.069) -0.140** (0.070) 0.055** (0.025) -0.043* (0.025) 4.90x10-6 (3.62x10-6) (0.041) -0.005 (0.033) -0.007 (0.017) 0.044** (0.022) -1.88x10-6 (2.41x10-6) SREMGRijkt (0.024) -0.008 (0.023) (0.040) 0.004 (0.032) -0.008 (0.016) 0.044** (0.022) -1.67x10-6 (2.43x10-6) (0.042) 0.003 (0.034) -0.010 (0.017) 0.049** (0.023) -1.85x10-6 (2.43x10-6) (0.027) 3.40x10-4 (0.015) -0.017 (0.016) 0.054*** (0.020) 6.70x10-8 (2.65x10-6) - - - USMWGijkt - - - 5.25x10-5*** (1.35x10-5) - -0.337 (0.378) 0.492*** (0.085) - Within R2 (Not presented for 2SLS regressions where it has little statistical meaning) Number of observations Number of skill groups 0.888 0.828 - - - 0.891 - - - 96 32 331 148 297 114 297 114 297 114 96 32 297 114 297 114 413 166 Weak instruments test: Smallest ‘maximal IV size’ bias rejected (Kleinbergen-Paap statistic) N/A N/A 20% (See [5]) 15% N/A 15% 15% 10% LREMGRijkt Standard errors clustered on the skill-group are reported in brackets. Significance at the 10%, 5% and 1% levels are denoted by *, **, *** respectively. The dependent variable in all but regression (4) is MWGijkt; the mean log Jamaican real wage in skill-group ijk at time t. EMGRijkt denotes emigrant share in skill-group ijk at time t, instrumented for using mean US real wages, USMWGijkt, in all 2SLS regressions, Tt a time dummy variable for time t, EDi education group, EXPj experience group, OCCk occupation group. The creation of cross-skill-group controls is explained in the appendix. BRAMWGijkt are Brazilian mean real wages as discussed in section 4biv), and SREMGRijkt and LREMGRijkt are the short and long run emigrant share variables discussed in section 4bv). Regression (4) presents the first stage fixed effects 2SLS results relating to regression (5) (thus the dependent variable is EMGRijkt). 20 iii) Benchmark fixed effects results: Columns [1] and [2] present the benchmark fixed effects results. The first column is included for comparison to the existing literature: skill groups are constructed on the basis of education and experience classifications only (four and eight groups respectively). It does, however, improve upon most papers through the inclusion of the cross-skill group controls. Stata does not support the use of time variant weights in fixed effects OLS regressions. As such, for both [1] and [2] the average number of remaining Jamaican workers used to calculate the dependent variable across all time periods for a skill group is used for the weight in every time period (again via the ‘aweights’ function). In [1], the coefficient of interest takes a value 0.426 and is significant at all reasonable significance levels with a p-value, corresponding to its t-statistic, of 0.000. A 10% increase in emigrant share is associated with a 4.26% increase in real wages of remaining workers of the same skill group. The own-skill-group effect however becomes insignificant both statistically and economically when classifying individuals by occupation too, producing many more skill groups, in [2] (note in these regressions I use four rather than eight experience groups). As suggested in Borjas (2007), the larger the number of skill groups used in classification, the larger measurement error is likely to be. Here I am not discussing non-random measurement error from undercount of illegal immigrants (discussed in section 5c)ii)), but rather measurement error that applies even to legal immigrants from the fact that the IPUMS US census data only captures a 5% sample of the population. Making the classical errors-invariance assumption that this type of measurement error is uncorrelated with the true emigrant share, it will lead to an attenuation bias in my estimate of θ1. This attenuation bias may be amplified in my fixed effects estimates if the persistence of true emigrant shares is considerably higher than the persistence of this measurement error. It is possible that the insignificance of θ1 in [2] stems from attenuation bias, which is suffered less in [1]. 21 iv) Understanding causality: An instrumental variables analysis: So far I have not addressed causality. In section 4b)i), I alluded to the possibility of reverse causation. Lower real wages, ceteris paribus, should encourage Jamaicans to emigrate, resulting in a downward bias in my estimates of θ1. I perform instrumental variables fixed effects regressions to try to use only exogenous variation in the emigrant share in identification of θ1 (that fraction of the variation in the emigrant share not caused by changes in Jamaican mean real wages). I use US average real wages across skill groups as my instrument for the emigrant share (USMWGijkt). If Jamaicans take into account the expected real wage they will earn in the US in their decision to emigrate, in line with the Harris-Todaro model, which in turn depends on the actual average real wage earned by existing workers of their skill group in the US, I expect this instrument to be highly relevant. This instrument was used in both Borjas (2007) and Gagnon (2011). Validity of the instrument: More questionable is the validity of my instrument. If changes in US average wages are directly correlated with changes in Jamaican average wages, over time within skill groups, other than through its effect on Jamaican emigration, my instrument is invalid. Part of the effect I attribute to emigration would rather capture this direct correlation. I control for one specific channel through which such direct correlation might exist: the existence of global technology shocks. A worldwide improvement in technology may increase the marginal product of labour and so real wages in both the US and Jamaica. I have income data (total rather than wage income) for one other country over these three time periods in the IPUMS dataset: Brazil. To the extent that correlation between Brazilian and Jamaican average income over time within skill groups captures those same global technology shocks that drive direct correlation between US and Jamaican average wages, by controlling for 22 Brazilian average income across skill groups I control for this potential abuse of my validity assumption. The fact that there is little migration between Jamaica and Brazil implies that correlation between their incomes is not driven by direct labour shifts between them, and so could plausibly be capturing common technology shocks. I should not, however, overstate the validity of my instrument. If there exist technology shocks common to the US and Jamaica but that do not impact Brazilian income, including the Brazilian control variable will not ensure the validity of my instrument. Further, one could think of other channels through which the validity assumption may be abused. An example is US import demand for Jamaican goods. Suppose US workers from better-educated skill-groups are those who can afford to visit Jamaica on holiday (the US Department of State website claims that in 2010 nearly 2 million Americans travelled to Jamaica). An increase in average wages for better-educated US workers could then translate through to higher wages for individuals in those same better-educated skill groups in Jamaica if they are employed in the tourist industry; leading to a direct positive correlation between US and Jamaican average wages over time within skill groups. In spite of this, my instrumental variables approach improves upon that attempted in the existing literature. Firstly, the two aforementioned papers to use US average wages take no additional steps to support its validity, unlike as I do. Secondly, Mishra (2007) rather uses the emigrant share across all experience groups in that given education group lagged one period as her instrument, as explained in section 2. In fixed effects regressions the consistency of our estimates depends on strict exogeneity of the independent variables. If the contemporaneous emigrant share is not strictly exogenous (due to reverse causality) then by definition neither is its lag; thus this instrument (based in part on the lagged emigrant share by skill group) will not be strictly exogenous either, and so her IV estimates are inconsistent. 23 IV results: The results for my 2SLS-FE regressions are presented in columns [3] and [5]. Other than instrumenting for the emigrant share and controlling for Brazilian average wages (BRAMWGijkt), I use the same specification as in [2]. Even if my instrument is both relevant and valid, IV estimates are only consistent, not unbiased. In small samples there may still exist large biases, and so I have chosen to use the education-experience-occupation classification as it provides me with a larger sample of observations. Stata supports the use of time variant weights when running fixed effects-2SLS regressions, as it was ‘easy enough to program the feature’ (Mark Schaffer, Statalist3). In [3] I use the same time invariant weights as in [2] for the purpose of comparison, whilst in [5] I use the time variant weights I had originally hoped to use in my fixed effects regressions. On this basis, the estimates in [5] are more efficient, although the estimates of θ1 are not dissimilar in economic or statistical significance. In [4] I present the first stage results relating to estimates in [5]. USMWGijkt has significant explanatory power over the emigrant share at all reasonable significance levels; t-statistic 3.88, p-value 0.000. US mean real wages appear to be a relevant instrument for the emigrant share, though I formally explore the identification of θ1 using this instrument below. Space precludes the inclusion of the other first stage results for fixed effects 2SLS regressions (with EMGRijkt as the first stage dependent variable), but USMWGijkt was significant at beyond a 1% significance level in all cases. In both [3] and [5] my FE-2SLS estimator of θ1 is identified. Since my errors are not i.i.d., at least due to serial correlation over time within skill groups, I must use the Kleinbergen-Paap rk LM test statistic to test for under-identification, a robust alternative to the Anderson and Cragg-Donald statistics. At even the 1% significance level in [5], I reject the null hypothesis of under-identification, given 3In an email sent to me 05/03/12 24 a p-value 0.0011, although in [3] it can only be rejected at a 2% significance level, p-value 0.0116. Baum, Schaffer and Stillman (2007) note that there has not yet been developed a test of weak instruments robust to non i.i.d. errors. However, given that the Stock-Yogo approach to testing weak instruments employs the same test statistic as under-identification tests (the difference being the null hypothesis tested and the critical values they have tabulated) the former authors suggest using the Kleinbergen-Paap rk Wald statistic (whose calculation is robust to non i.i.d. errors), with the Stock-Yogo critical values to test for the presence of weak instruments. Since these critical values have not been created specifically for this test, conclusions must be drawn cautiously. In [5] using the Kleinbergen-Paap rk Wald statistic I reject the null hypothesis that the instrument is sufficiently weak that Wald tests of the IV estimates have a rejection rate 15% or above (as opposed to the true rate of 5%; the ‘maximal IV size’ test), given a statistic of 15.04 compared to a Stock-Yogo 15% maximal IV size critical value of 8.96. However, in [3] I can only reject at a 20% maximal IV size level. Nonetheless, I tentatively conclude that my 2SLS estimates do not suffer severe problems from weak instruments. According to the estimate of θ1 in [5], a 10% increase in emigrant share causes a 4.34% increase in real wages of remaining workers of the same skill group. This coefficient is statistically significant at a 1% significance level, with p-value 0.001. Not only is the estimate now statistically significant, it has also increased significantly economically from the fixed effects estimate in [2], consistent with the instrument removing a downward bias created by the previously described reverse causation. Thus even if the attenuation bias suffered in [2] has not been removed, given I have not tried to argue that my instrument is uncorrelated with the measurement error in the emigrant share, the removal of the bias from reverse causation is sufficient to reveal a positive and significant own-skill-group effect. 25 The coefficient on cross-skill control 1 is negative and significant at the 5% level. The cross-skill controls have been constructed hierarchically, cross-skill control 1 is always the emigrant share of the lowest educated of the ‘other three’ education groups controlled for (cross-skill control 3 is always the highest). The negative sign suggests emigration of relatively low educated workers is detrimental to others’ real wages. To explore whether a negative cross-skill effect only operates when that emigrating worker is less educated than you I created three dummy variables; one that takes value 1 when EMGRijkt is the emigrant share for primary completed, one for secondary completed, and one university completed. I interact each dummy variable with each cross-skill control. If complementarities only exist with lower educated workers, the interaction terms should only be significant when indicating the emigrant share of a lower educated group. This was not the case however (results not reported). Given this, and given the sign and significance of each cross-skill control is less robust across specifications, I simply conclude there is suggestive evidence that cross-skill effects do exist, so that failing to control for them would create a bias in the estimate of θ1 were they correlated with EMGRijkt. BRAMWGijkt is insignificant (p-value 0.436). To the extent that there exists direct correlation between US and Jamaican average wages over time within skill groups due to common technology shocks, those technology shocks do not seem to have driven Brazilian income in a correlated way. Global technology shocks and US import demand as previously described would likely drive positive correlation between US and Jamaican average wages. If these were the main sources of concern about the validity of the instrument, the estimate of θ1 in [5] perhaps constitutes an upper bound, with the estimate in [2] a lower bound, for the true effect of emigration (ignoring all other econometric problems). v) Short run versus long run: The analysis so far calculates the emigrant share based on the total stock of workers in the US at that point in time. It assumes the effect of emigration is permanent; the real wage effect on remaining workers is the same whether an 26 individual left, say, 2 or 20 years ago. For the reasons expressed in section 3v), one might expect the effects of emigration to change over time. I discussed that the own-skill group effect may be supressed in the long run. I therefore divide the emigrant share into a ‘short run’ and ‘long run’ effect, similar to Mishra (2007), although using a different definition of short and long run. I use a variable ‘mgrate5’ from the IPUMS dataset, which expresses where an individual resided 5 years before the census date. I separate those Jamaican emigrants who lived ‘abroad’ five years ago from those who lived in the US. I treat the former as capturing migrants who emigrated in the last five years, and the latter as having moved more than 5 years ago. The measure of recent migrants is not perfect given that it also captures individuals who left Jamaica for a country other than the US many years ago, before moving to the US within the last five years. Nonetheless, it presents the best approximation available using the census data. Failing the existence of two separate instruments for the short and long run emigrant share (simultaneously using current and lagged US mean real wages led to under-identified models), I was forced to return to my simple fixed effects estimates. I look to see whether the significant own skill group effect in [1] embodies both a significant short and long run effect of emigration. By contrast to the theory, in [6], I apparently find a significant long-run effect at all reasonable significance levels (p-value 0.000), but the short-run effect is insignificant, and even switches sign. However, this is perhaps to be expected econometrically. Whilst a decrease in mean wages today may have caused an increase in emigration from a skill group within the last five years, it will not have caused emigration a decade or more previously. That is to say reverse causation will bias the coefficient on the short run emigrant share downwards, but not the long run. The absence of two or more instruments precludes further analysis. c) Tests of robustness i) Sample selection bias: 27 Jamaican emigration outside of the US: Interpreted as the own-skill group effect of Jamaican emigration to all destination countries, my estimates may suffer from sample selection bias given I only observe emigrants to the US. Given I estimate using fixed effects, time invariant influences on selection to the US against other destinations will not bias my estimates, yet unobserved time variant influences may if correlated with the disturbance term. Since my estimates are identified from within variation in emigrant shares, it is reassuring that according to the aforementioned World Bank panel data set 85.6% of the total increase in Jamaican emigration to the six OECD countries (including US, UK and Canada) 1980-90, and 91.0% of that 19902000, occurred in the US. Whilst my estimates are plausibly identified from the majority of total within variation in emigrant shares, I should retain some caution in interpreting results as representative of the effects of emigration outside of the US. Unobserved ability of emigrants: An alternative and potentially severe sample selection bias stems from the fact that I do not observe the inherent ‘ability’ of workers. If less able workers emigrate from Jamaica, the average ability of remaining workers increases, and thus the average marginal product of labour (and so real wages) will increase, quite aside from any effect of emigration on real wages. A priori it is not possible to sign sample selection bias, since if more able workers emigrate then by reverse argument my estimates are downward biased. Fundamentally, the sign of this sample selection bias depends on where in the conditional (on individual characteristics controlled for in my regression: education, experience and occupation) wage distribution in Jamaica emigrants have come from. This is the same sample selection problem discussed by Clemens, Montenegro and Pritchett (2010) (CMP). Ceteris paribus one would expect workers in the lower tail of the Jamaican conditional wage distribution to be more likely to emigrate (i.e. ‘less able’ individuals). However, the ability and desire to emigrate 28 may depend on a host of unobservable characteristics that also influence the wages one obtains in the Jamaican conditional wage distribution. As an example, CMP comment that the networks one has in the US will influence one’s ability to emigrate there. Extending their discussion, consider the unobserved ability: charisma. An individual who has this ability is both more likely to have a lot of friends, and so better networks of individuals who have already migrated to the US, but also more likely to gain promotion in their workplace and so come from the higher tail of the Jamaican conditional wage distribution. This would imply positive selection into emigration. CMP provide empirical evidence to infer the likely type of selection into emigration. The external validity of results may be limited given the barriers (language, distance etc) workers must overcome to emigrate from one country may be very different to another, and so the unobserved characteristics that become important differ too. In this light perhaps the most interesting evidence from CMP comes from emigration of workers from Costa Rica, Dominican Republic and Haiti, countries geographically similar to Jamaica, to the US. Comparing the conditional wage distribution of migrants’ pre-migration wage to the conditional wage distribution of non-migrants, they estimate that the average migrant is selected from the 50th-60th percentile of the distribution of non-migrants, consistent with positive selection. Whilst positive selection into emigration seems more plausible empirically, to understand the robustness of my results I consider what degree of negative selection would be necessary to completely eradicate the positive and significant effect of emigration on real wages estimated in my preferred 2SLS-FE results in [5]. This is a more desirable and transparent approach to that taken by Mishra (2007), Gagnon (2011) and Bouton (2011). They strangely estimate the ownskill-group effect for a regional sub-sample of workers who are unlikely to emigrate on the grounds that if virtually no one emigrates there cannot be any sample selection bias! 29 Suppose 10% of every skill group emigrates, and the only individuals to emigrate are of ‘low ability’. Before emigration, the average marginal product of labour, if high ability workers have MPL = 100 and low ability have MPL = x, is: y = (10x + (90*100))/100 After emigration, the average MPL is 100 (only high ability stay behind). The percentage increase in the average MPL due to sample selection is therefore = ((100-y)/y)*100. That same emigration of 10% of the initial workforce would imply in my regressions an increase in the emigrant share variable of 10/90. From [5], the associated percentage increase in average real wages is = ((10/90)*100)*0.43392 = 4.821%. Therefore for the entire positive and significant effect of emigration to be removed by sample selection bias: ((100-y)/y)*100 = 4.821. Solving for y, before substituting back into x, gives x = 54.0. This simple simulation exercise suggests that for the entire positive and significant effect of emigration on real wages in my preferred estimates to be removed, I would have to assume that, controlling for education, experience and occupation of an individual (where 10% of the workforce emigrates and they are exclusively the low ability workers), low ability workers on average have a marginal product of labour only just over half that of remaining workers. An extreme level of negative selection into emigration would be required to entirely refute my results. ii) Measurement error: The main source of measurement error I must deal with is the undercount of Jamaican emigrants due to the existence of Jamaican-born illegal immigrants living in the US. For this form of measurement error I cannot make the classical 30 errors-in-variance assumption since the extent of the undercount of illegal immigrants is likely correlated with the true level of emigration in a skill group. This undercount of Jamaican emigrants is an omitted variable. Mishra (2007) assumes that all illegal immigrants are of her lowest educated category, and that there is an undercount of 30%. She performs regressions adjusting emigrant shares accordingly for those skill groups. Given that I control for all time invariant variables, perhaps a greater concern is if there is a significant change in the extent of undercounting over time. I suppose rather that all ‘less than primary education completed’ and ‘primary completed’ skill groups undercount emigrants by 30% in the first period, but, due to an improving ability of the authorities to track and deport illegal immigrants, this undercount is only 20% and 10% in the second and third period respectively in [7]. I then, by contrast, assume that the problem of illegal immigration worsens, such that the undercount increases to 40% and then 50% in the second and third periods in [8]. In all regressions I have used the 2SLS-FE approach from my preferred regression [5]. The estimated wage elasticity changes very little in either case, reducing to 0.426 in the former and increasing to 0.456 in the latter. iii) Other robustness checks: Finally, due to the fairly arbitrary nature of the experience classification, I check the robustness of my results in [5] to the use of the eight-group rather than the four-group experience classification, in [9]. Again the estimated wage elasticity changes very little, increasing to 0.461 and remaining significant at all reasonable significance levels, p-value 0.000. 5. Conclusion To summarise, I find significant evidence of a positive own-skill-group effect of emigration on Jamaican real wages, plausibly robust to concerns about sample selection and the undercount of Jamaican illegal immigrants in the US. My 31 preferred estimates suggest that over the period 1980-2000, a 10% emigrationinduced decrease in the Jamaican labour force in a skill group would cause a 4.34% increase in the real wages of remaining workers of the same skill group. My estimates are qualitatively and quantitatively consistent with the existing literature. Considering regressions with the most comparable sample and specification to those in my essay, own-skill-group wage elasticities have been estimated in the range 0.21 (Borjas (2007)) to 0.56 (Aydemir and Borjas (2006)). These are fixed effects estimates that do not deal with causality and classify workers in terms of education and experience only. They are most comparable to my estimate from [1] of 0.342, which falls within the range. It is interesting that authors across a range of developing countries find evidence of a similar several percentage point own-skill-group effect of a 10% emigrationinduced decrease in the labour supply, in spite of the fact that determinants of the emigration decision (and thus sample selection issues), and the level of source country unemployment, are very context-specific. Implications for poverty reduction and policy: As I have stressed throughout, conclusions about the overall effects of emigration cannot be drawn from an understanding of the own-skill-group effects alone. Nonetheless, a brief thought experiment demonstrates the likely implications of the economic magnitude of the estimates for poverty reduction for the wage-earning Jamaican poor. Suppose there is emigration from a single skill-group only, such that no cross-skill group effects influence wages in that skill-group. Suppose 22.1% of the labour force in that skill-group emigrates (the median decadal change in emigrant share amongst those in which the emigrant share increased). From estimates in [5], this will produce an increase in real wages of remaining workers by 9.6%. A relatively poor Jamaican worker earning the equivalent of US$1,000 per annum could expect to see an increase in his real wage to US$1,096. Abstracting from any other benefits of emigration for the workers who stay behind (such as remittance flows), whilst emigration may pull the emigrants themselves out of poverty (according to the estimates of CMP 32 (2010) mentioned in section 1) it appears unlikely, alone, to produce real wage increases of an order of magnitude sufficient to pull remaining workers out of poverty. I do however caution against drawing policy conclusions from this analysis alone. The own-skill-group effect estimated in this extended essay could be combined in the future with attempts to gauge the size of cross-skill-group effects in Jamaica to identify the remaining workers likely to suffer or gain from the actual pattern of Jamaican emigration. Policy may then in turn aim to assist those groups who suffer. Currently however, no such attempts exist. 33 Appendix Table 2: Data description Variable name MWGijkt Description Data source Procedures applied to raw data Mean log real wage of remaining workers in Jamaica IPUMS International: Jamaica 1982, 1991, 2001; IMF International Financial Statistics EMGRijkt Emigrant share Cross-skill controls (1, 2 and 3) Cross-skill group emigrant shares IPUMS International: Jamaica 1982, 1991, 2001; US 1980, 1990, 2000 IPUMS International: Jamaica 1982, 1991, 2001; US 1980, 1990, 2000 SREMGRijkt Short-run emigrant share I converted nominal annual wage income from all sources of employment (given by the ‘incwage’ variable in IPUMS) into real terms using the Jamaican CPI in the annual IFS series (dates 1982, 1991, 2001), before taking its natural log. I then calculated the mean log real wage across all workers for whom this variable was not missing in the IPUMS data set in each skill group. To do so I multiplied each log real wage by the ‘wtper’ variable (which gives the number of workers in the population represented by that sample point) and summed this across all workers in the skill group (to give total wage income in the skill group), before dividing by the number of workers in the skill group. I took the number of Jamaican born workers in the US using the ‘wtper’ variable in the US census data. I then divided this, for each skill group, by the total number of individuals (employed or unemployed) in Jamaica in that skill group (using the Jamaican wtper variable). The three cross-skill control variables are simply the emigrant share in each of the three skill groups with the same experience and occupation but different level of education to the skill group of the dependent variable. Cross-skill group control 1 is always the lowest of the ‘other three’ education groups, cross-skill control 2 the second lowest, and cross-skill 3 the highest. Identical to EMGRijkt, but counting only those Jamaican born workers in the US with a value ‘abroad’ for the ‘mgrate5’ variable in the US census in the numerator. LREMGRijkt Long-run emigrant share BRAMWGijkt Brazilian mean real wage US MWGijkt US mean real wage (used as the instrument for EMGRijkt) IPUMS International: Jamaica 1982, 1991, 2001; US 1980, 1990, 2000 IPUMS International: Jamaica 1982, 1991, 2001; US 1980, 1990, 2000 IPUMS International: Brazil 1980, 1991, 2000; IMF International Financial Statistics IPUMS International: US 1980, 1990, 2000; IMF International Financial Statistics Identical to EMGRijkt, but counting only those Jamaican born workers in the US with a value not equal to ‘abroad’ for the ‘mgrate5’ variable in the US census in the numerator. I first converted total nominal monthly income as given by ‘inctot’ in the Brazilian census into a single currency: the Cruzeiro in use in Brazil in 1980 (IPUMS reports figures in terms of the currency in use at that census date). I then deflated nominal values using the Brazilian CPI in the annual IFS series. Finally I took the mean Brazilian real wage across skill groups in the same way as for the US above. Identical procedure to MWGijkt, deflating using annual IFS series ‘US CPI all items city average’. 34 References Aydemir. A and Borjas. G. 2006. ‘A Comparative Analysis of the Labour Market Impact of International Migration: Canada, Mexico and the United States’, NBER Working Paper No. 12327. Baum. C, Schaffer. M and Stillman. S. 2007. ‘Enhanced Routines for Instrumental Variables/GMM Estimation and Testing’, Boston College Working Papers in Economics No. 667. Bertrand. M, Duflo. E and Mullainathan. S. 2004. ‘How Much Should We Trust Difference-in-Difference Estimates?’, The Quarterly Journal of Economics, Vol. 119, No. 1, pp. 249-275. Borjas. G. 2003. ‘The Labour Demand Curve is Downward Sloping: Re-examining the Impact of Immigration on the Labour Market’, The Quarterly Journal of Economics, Vol. 118, No. 4, pp. 1335-1374 Borjas. G. 2007. ‘Labour Outflows and Labour Inflows in Puerto Rico’, NBER Working Paper No. 13669. Bouton. L, Paul. S and Tiongson. E. 2011. ‘The Impact of Emigration on Source Country Wages: Evidence from the Republic of Moldova’, World Bank Policy Research Working Paper No. 5764. Boyer. G, Hatton. T and O’Rourke. K. 1993. ‘The Impact of Emigration on Real Wages in Ireland 1850-1914’, CEPR Discussion Paper No. 854. Clemens. M. 2011. ‘Economics and Emigration: Trillion-Dollar Bills on the Sidewalk?’, Centre for Global Development, Working Paper Number 264. 35 Clemens. M, Montenegro. C and Pritchett. L. 2008. ‘The Place Premium: Wage Differences for Identical Workers across the U.S. Border’, Centre for Global Development, Working Paper Number 148. Dustmann. C and Glitz. A. 2011. ‘How Do Industries and Firms Respond to Changes in Local Labour Supply?’, CREAM Discussion Paper No. 18/11 Elsner. B. 2011. ‘Emigration and Wages: The EU Enlargement Experiment’, IZA Discussion Paper No. 6111 Gagnon. J. 2011. ‘”Stay With Us?” The Impact of Emigration on Wages in Honduras’, OECD Development Centre Working Paper No. 300. Hanson. G. 2005. ‘Emigration, Labour Supply and Earnings in Mexico’, NBER Working Paper No. 11412. International Monetary Foundation. ‘International Financial Statistics’. Katseli. L, Lucas. R and Xenogiani. T. 2006. ‘Effects of Migration on Sending Countries: What Do We Know?’, OECD Development Centre Working Paper No. 250. Kim. N. 2007. ‘The Impact of Remittances on Labour Supply: The Case of Jamaica’, World Bank Policy Research Working Paper No. 4120. Lucas. R. 1987. ‘Emigration to South Africa’s Mines’, The American Economic Review, Vol. 77, No.3, pp. 313-330. Minnesota Population Centre. ‘Integrated Public Use Microdata Series International’, University of Minnesota. Mishra. P. 2007. ‘Emigration and Wages in Source Countries: Evidence from Mexico’, Journal of Development Economics, Vol. 82, pp.180-199. 36 Schiff. M and Sjoblom. M.C. The World Bank Trade Team Development Research Group. ‘Panel Data on International Migration 1975-2000’: http://econ.worldbank.org/WBSITE/EXTERNAL/EXTDEC/EXTRESEARCH/0,,co ntentMDK:21866422~pagePK:64214825~piPK:64214943~theSitePK:469382,0 0.html U.S. Department of State. 2012. ‘Background Note: Jamaica’: http://www.state.gov/r/pa/ei/bgn/2032.htm Wooldridge. J. 2002. ‘Econometric Analysis of Cross Section and Panel Data’, MIT Press, Cambridge, Massachusetts, United States. The World Bank. ‘Bilateral Migration Database 1960-2000’. The World Bank. ‘World Development Indicators’. 37