Distributional Effect of Import Shocks on the British Local Labour Markets

Distributional Effect of Import Shocks on the British Local Labour Markets Anwar S. Adem Lancaster University August 2018 Preliminary Version Abstract In this paper, we investigate the causal effect of import shocks on wage inequality using individual-level data from the United Kingdom in the period 1997-2010. Methodologically, we exploit regional variation in initial industrial structure and concentration for identification and apply a group IV quantile approach. The data allow us to control for individual characteristics and this proves important to discover that an increase in import competition contributes for the rise in wage inequality. While the import shock negatively and significantly affecting workers at the bottom of the wage distribution, its effect on the median and upper wage groups is not significant. Lastly, the analysis on subsample finds that the negative effects are concentrated on those low-middle male, full time and manufacturing workers. Keywords: Wage inequality, import competition, local labour market, quantile regression, instrumental variables JEL: F14, F16, J31, R23 I am grateful to Maria Navarro Paniagua and Maurizio Zanardi for their guidance and supervision. I also thank Jean-Francois Maystadt for his helpful comments. Financial support from the Economic and Social Research Council (ESRC) under ESRC NWDTC Standard Studentship is greatly acknowledged. The paper also benefited participant comments at NWSSDTP conference at the University of Liverpool. All errors are mine. Contact details: Department of Economics, Lancaster University, Bailrigg, LA1 4YX, Lancaster, United Kingdom E-mail: a.adem@lancaster.ac.uk

1. Introduction In the past three decades, the labour market of most developed countries has been characterized by the decline in the share of manufacturing workers and an increase in wage inequality (Autor, Dorn, & Hanson, 2016). In this regard, the rise in wage inequality is particularly well documented, for instance, from the mid-1980 onwards: inequality has increased in 17 out of 22 OECD member states for which data is available (OECD, 2011). Here, the often mentioned causes are skilled-biased technological change, institutional set-ups and globalization, that is, the rise in international trade and off-shoring (Acemoglu & Autor, 2011; Piketty, 2014). And within the globalization channel, the rise of China as an exporting power has received much attention. In this paper, we investigate the causal link between import competition and wage inequality using worker-level data from the United Kingdom for the years 1997-2010. The UK s labour market represents an interesting case to investigate the effect of trade shock on wage inequality, as it has experienced both the rise in wage inequality and a substantial increase in import competition over the past three decades. Figure 1 illustrates these two facts by showing the trend in the import values and wage inequality. Unlike other developed European countries, the UK is a country with a high degree of inequalities especially among its regions: the inter-regional inequality of the UK is above the OECD average. In fact, it is the only European country with relatively large regions 1 in all five quantiles of the EU GDP per-capita distribution (Gibbons, Overman, & Pelkonen, 2010; McCann, 2016; Arellano & Bonhomme, 2017), with regional divergence being the phenomena that began to accelerate in the 1990s (McCann, 2016). Instead, similar to most other developed countries, its import from China have grown exponentially following China s accession to the World Trade Organisation (WTO) in 2001. Given that regions within a country differ in their industrial structure and concentration of activities, their exposure to trade shocks is also likely to differ. instance, regions specialized in textiles would be affected more by increased import competition from low wage countries than regions specialized in auto-mobile manufacturing. Thus, this regional variation in the degree of exposure to trade shocks is commonly exploited to identify the effect of import surges and answer research questions related to the adjustment of local labour markets to trade shocks, with 1 There were 37 NUTS-2 regions in the UK according to NUTS version 2003 and 2010, and there are 40 NUTS-2 regions according to 2013 version For 1

useful dis-aggregations by sectors. This paper exploits similar identification strategy to investigate how import shocks contribute to the rise in wage inequality. In the search for a clean identification strategy to provide causal evidence on the role of trade shocks on the local labour market, we also follow the literature and use China s accession to the WTO. The underlying assumption is that this policy change represents a quasi-natural experiment, as it can be assumed to be unanticipated from the point of view of UK regions in general and UK firms within the regions in particular. Therefore, it is an exogenous shock for all regions regardless of their industrial composition. Figure 1. Trend in Total UK Import Value from China and Inequality Total Import Value in Billion of Pounds 0 5 10 15 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 Year (a) Trend in Import value (b) Trend in Wage Inequality Notes: (a) Author calculation using WITS dataset from the World Bank. Values are in billion of 1987 pounds. (b) Source: Resolution Foundation Thanks to the increased availability of micro datasets (i.e., firm-level data, worker-level data and matched employer-employee datasets) researchers have zoomed their attention to investigate the local or regional labour market effect of trade liberalization (Kovak, 2013; Autor, Dorn, & Hanson, 2013; Dix-Carneiro & Kovak, 2015). This paper is also related to this strand of literature by examining the local labour market effects of trade shock using a representative sample of worker-level data. This paper is also related to the growing literature that examines the differential effects of trade shocks on labour markets outcomes in general and wage inequality in particular. To this end, we use the group IV quantile approach developed by Chetverikov, Larsen, and Palmer (2016). Moreover, to correct for the endogeneity problem associated with our variable of interest, we follow the empirical approach advocated by Autor et al. (2013) (ADH hereafter). These approaches help us to 2

identify the effect of import competition on wages of workers at different quantile levels, and by comparing these effects we can analyse the effect on trade liberalization for different groups. In general, our analysis contributes to the literature in two ways. First, methodologically, we extend the analysis of IV quantile by controlling for individual characteristics such as age, sex and occupation. Our findings demonstrate that the inclusion of these controls is of great importance: the inclusion of these covariates increases the precision of the estimated coefficients at every quantile level leading to different conclusions. Specifically, high import exposure contributes to the rise in wage inequality as it negatively affect workers at the bottom of the wage distribution while leaving those at the upper-middle and top unaffected. The effect is particularly significant for those between 20 th to 30 th quantiles of the wage distribution.moreover, further disaggregation of workers into sub groups, we find that the effect of import competition is mainly concentrated on the male, full time and manufacturing workers who are at the lower part of the wage distribution. Second, from an empirical perspective, we answer the research question for the UK, which represents interesting country for the reasons mentioned above. Furthermore, in addressing both qualitative and quantitative effects, we decompose the wage effect into hourly wage and total hours worked. The remainder of this paper is organized as follows. Section II discusses related and recent developments in the literature. In section III, we describe the data sources and present descriptive statistics. Section IV explains the methodological framework. While the main results are presented with section V. Section VI concludes. 2. Related Literature The causal effect of globalization on labour market outcomes has been an important research topic and has featured on the policy agenda. This is because while gains from increased imports are spread across the economy, losses are concentrated among the direct competitors, and hence the overall effect of import competition on employment and wages of the latter group is non-trivial (Greenland & Lopresti, 2016). Therefore, studies in this area are not only trying to answer questions related to the welfare effect of globalization, that is, who gains and who loses from globalization, but also how to properly implement distributive policies. 3

Meanwhile, researchers in the past have attributed the observed reduction in manufacturing workers and the rise in wage inequality in most developed countries to, inter alia, skill-biased technological progress, institutional set ups, and trade. However, high emphasis on the former two overshadowed the attention given to the effect of trade shocks. Due to the limited evidence on local labour market effects of trade shocks, the causal link between the two has remained ambiguous for long (Krugman, 2008; Bloom, Draca, & Van Reenen, 2016). The previous literature in the area also concentrated on the effect of trade on economy wide outcomes. The use of microeconomic data and different features of recent trade relations distinguish recent studies from their previous counter parts. Micro data availability allows researchers to investigate the causal effect of trade shocks on local labour markets at more disaggregated levels. In this regard, one strand of the literature uses reduced-form analysis to study the impact of trade shocks on welfare, employment, and wages at different levels of disaggregation. Similar to most previous studies, these studies find that trade-related demand shocks cause different effects on different sub-economies or groups. However, unlike previous studies whose analysis was limited to the effects of trade shocks on different owners of resources, that is, capital and labour, recent studies analyse trade shocks on different regions within a country (Topalova, 2010; McCaig, 2011; Kovak, 2013), on different occupations (Acemoglu & Autor, 2011; Ritter, 2014; Peri & Sparber, 2009), on different industries (Revenga, 1992; Attanasio, Goldberg, & Pavcnik, 2004), on different occupations and industries (Utar, 2016; Artuç & McLaren, 2015), and even among different age groups (Artuç, 2012). In her pioneering work, Topalova (2010) investigates the effects of variation in tariff reductions on poverty levels of districts in India. After constructing tariff reduction intensities for each district based on variation in sectoral composition, she finds that regions which are more exposed to trade, through higher tariff reduction, experience slower decline in poverty and consumption. She also shows that the effect is severe for less mobile groups of workers such as those at the bottom of the income distribution. Likewise, Kovak (2013) exploits the regional variation in exposure to trade shocks to analyse local labour market effects in Brazil. His findings show the presence of negative, location-specific effects from trade shocks on wages and employment. Both of these studies use employment-weighted average tariff as a measure of exposure. 4

Similarly, Hakobyan and McLaren (2016) analyse the effect of tariff reduction in the U.S. following NAFTA on a wage of workers at local labour market level between 1990 and 2000. Their reduced-form analysis finds a large effect of NAFTA on wages of regions and industries that experienced a larger reduction of tariffs. Chiquiar (2008) also analyses the effect of NAFTA on Mexico s local labour market. He finds an increase in wage inequality and overall wage levels and a decline in skill premium for highly exposed regions. The second strand of literature uses structural models to investigate the dynamics of labour markets following trade shocks. Structural models combine economic theories and statistical methods to develop mathematical relationships between economic variables. Beyond offering insights into the causal link of exogenous and endogenous variables, these studies allow researchers to model the underlying mechanism through which causal relations operate. Therefore, they help us to answer a multitude of research questions (Reiss & Wolak, 2007). Unlike reduced-form studies, structural studies find mixed evidence on the effects of trade shocks on labour market outcomes, particularly on wage inequalities. On the one hand, there are studies that find a considerable impact of trade liberalization on wage inequality. For example, after structurally estimating a heterogeneous trade model with imperfect labour market of search and matching, Helpman, Itskhoki, Muendler, and Redding (2012) find a significant effect of Brazilian trade liberalization on wage dispersion for the period spanning 1986-1998. Similarly, Dix-Carneiro and Kovak (2015) use Brazilian data, to show the presence of lag in adjustment following trade shocks and cost of mobility. Egger, Egger, and Kreickemeier (2013) develop and estimate a structural model that combines heterogeneous firm model and worker with fair-wage preferences, and find a non-negligible impact of openness on the wage inequality using data from France and Balkan countries. On the other hand, Coşar, Guner, and Tybout (2016) find no evidence on the impact of trade liberalization, per se, on wage inequality after analysing Colombian trade and labour market reform. Likewise, Prat, Felbermayr, Impullitti, et al. (2016) find no evidence to say trade openness of Germany is the cause for the increase in wage inequality. In addition to using different approaches and levels of disaggregation, studies also differ in their local labour marker outcome of interest. Some studies analyse the effect on employment level; other studies investigate the effect on wage and wage inequality; others aim at analysing the effect of import competition on productivity, 5

innovation and R&D related investments; and still others are concerned with the welfare implications of trade shocks. The different characteristic features of recent South-North trade patterns also demand reinvestigation of the causal effects of trade shocks on the rise of wage inequality and decline in manufacturing employment (Helpman et al., 2012; Sampson, 2016; Prat et al., 2016). The recent increase in trade relations demands re-investigation since it might have different implication on the impact of trade shocks on labour market outcomes (Bloom et al., 2016). In this regard, a growing literature uses China s accession to WTO as an identification strategy. The seminal paper by Autor et al. (2013) analyses the distributional consequences of import competition from China on U.S commuting zones. They use the presence of variation in industrial structure and specialization among these zones to exploit variations in trade exposure. The authors find that more exposed regions experienced lower wages, reduced employment prospects, increased transfer payments from federal and state programs. Another study for the US include a work by Pierce and Schott (2016) which uses China s grant of Permanent Normal Trade Relation (PNTR) status to exploit the impact of import competition on U.S employment and find similar result as ADH. However, recent studies by Feenstra, Ma, and Xu (2017) and Wang, Wei, Yu, and Zhu (2018) find contrary result.... Other recent studies which use a similar methodology include a study by Balsvik, Jensen, and Salvanes (2015) who highlight the negative effect of Chinese import competition on employment, particularly, on those low-skilled ones using Norwegian data. Dauth, Findeisen, and Suedekum (2014, 2016) analyse the effect of the rise of import from China and Eastern Europe on regional labour markets of Germany from 1988 to 2008 and 1990 to 2010. Mendez (2015) finds that highly exposed regions to import competition from China experience a larger reduction in the share of manufacturing employment and greater mobility of workers using Mexican data. Our paper is also related to this strand of literature by examining the causal effects of trade shocks: trade liberalization, expansion of exporters, and lower trade cost, on wage inequality. Methodologically, a study by Han, Liu, and Zhang (2012), which analyses the causal effect of Chinese accession to WTO on wage inequality of high exposed and low exposed regions of urban China, is closely related to ours. By analysing the 6

presence of significant changes at the 90 th and 10 th quantiles, they show how import competition exacerbates or cushions wage inequality. However, the present paper uses a recently developed measure of regional import exposure, and the analysis focuses on the United Kingdom. 3. Data Description In order to answer our research questions around wage inequity and trade, a wealth of data are required. In this end, we used four data source: Annual Survey of Hours and Earnings (ASHE), Annual Business Inquiry (ABI) and Business register and employment survey (BRES), UN Comtrade, and OECD regional statistics and indicators. Further detail is available in Appendix A. Data on individual workers and their characteristics come from the Annual Survey of Hours and Earnings (ASHE) dataset from the UK data service. It provides us with a 1% sample of employees from National-Insurance records for the years 1997 to 2010, which determines our analysis period. Since the data is reported by employers from HM Revenue and Customs PAYE of employees, the information is accurate. Most importantly, this dataset includes variables on both employee and employer characteristics. Variables referring to employee characteristics include: wages, hours of work, occupation, gender, and full time status. From the employer side, characteristics such as employment size, industry classification and legal status (e.g. public company) are included in the dataset. In addition, the dataset is representative of 128 NUTS3 regions and industries (ONS, 2017a). Our identification strategy involves changes in labour market outcomes before and after China s accession to WTO in December 2001. Thus, our analysis will exploite the changes across the two periods: 1997 to 2002 and 2002 to 2010, with the changes weighted to represent decadal changes. 2 Table 1 reports descriptive statistics of the main variables from ASHE for three years, that is, 1997, 2002 and 2010. As we will discuss in the section below, the construction of change in regional import exposure requires weighting the change in import by a number of manufacturing employment. Therefore, we use data on the regions number of manufacturing 2 Following ADH, we convert this changes into their decadal equivalence changes by multiplying changes by 10/5 and 10/8 for the year 2002 and 2010 respectively. 7

workers from Annual Business Inquiry (ABI) and Business Register and Employment Survey (BRES) of official labour market statistics Nomis(ONS, 2017b). These data is available at the 3-digit level of SIC 2003 which determines the level of industrial disaggregation. Furthermore, import data are derived from the World Integrated Trade Solution (WITS)(WITS-UNSD, 2017). This database imports data from UN Comtrade, under different industries classified with the 3-digit level of NACE Rev.1 because at this level it is identical to the SIC 2003 classification used in ABI and BRES. Table 1. Descriptive Statistics Q25 Q50 Q75 Mean Std.err Min Max Year 1997 Real average gross weekly earnings 195.43 326.73 495.25 379.29 322.05 0 >8,000 Real hourly earnings 5.76 8.40 12.70 10.17 8.68 0 >220 Average total paid hours worked 27.5 37 40 32.98 14.57 0 >125 Male 0 1 1 0.52 0.49 0 1 Occupation (isco88) 2 4 7 4.5 2.51 1 9 Full time 0 1 1 0.72 0.45 0 1 Manufacturing 0 0 0 0.21 0.40 0 1 Age 29 38 48 38.77 11.67 16 64 Observations 148,759 Year 2002 Real average gross weekly earnings 229.70 373.12 572.08 449.59 424.62 0 >12,500 Real hourly earnings 7.01 9.89 15.11 12.68 11.34 0 >360 Average total paid hours worked 31.25 37 40 34.64 10.97 <1 >100 Male 0 1 1 0.51 0.49 0 1 Occupation (isco88) 3 4 7 4.63 2.57 1 9 Full time 1 1 1 0.76 0.43 0 1 Manufacturing 0 0 0 0.16 0.37 0 1 Age 30 39 49 39.55 11.80 16 64 Observations 151,472 Year 2010 Real average gross weekly earnings 230.52 379.62 588.86 461.69 407.88 0 >9,150 Real hourly earnings 7.5 10.45 16.10 13.54 11.70 0 >380 Average total paid hours worked 27.5 36.99 39.79 33.13 11.74 0 >110 Male 0 0 1 0.50 0.50 0 1 Occupation (isco88) 2 4 5 4.46 2.53 1 9 Full time 0 1 1 0.71 0.45 0 1 Manufacturing 0 0 0 0.10 0.30 0 1 Age 30 40 50 40.05 12.21 16 64 Observations 165,544 Notes: Real monetary units in 2010 pounds; some of the minimum and maximum values are suppressed for disclosure avoidance; and a full list and explanation of the variables are provided in the appendix. Using information from the above two sources, we can calculate the change in import exposure to China for each of the 128 NUTS 3 region of Britain for which we also have representative data from ASHE. Thus, these NUTS3 regions are considered as local labour market in Britain. 3 3 Other studies use travel to work areas as local labour markets in the UK. We do not consider travel to work since we do not have data on regional characteristics at that level of classification. NUTS 3 regions are slightly bigger but not too big to be considered as an alternative. Assuming the reduction in commuting cost and increasing number of workers who commute to work, considering NUTS 3 as the local labour market is reasonable. 8

In particular, we use the region s share of employment in industry j and the change in imports per number of workers to calculate change in import exposure. That is, we sum changes in import values per regional employment across 93 industries and weighting them by regional share of employment in each industry: IP W UK rt = j Emp rjt Emp jt IMP jt Emp rt, (1) where Emp rjt represent start of period regional employment in industry j, Emp jt stands for start of period the total number of employment in industry j in Britain, Emp rt is start of period number of employment in region r, and IMP jt stands for the change in value of industry j s import by the UK from China (in 1000s of pounds). Figure 2. Decadal Change in Regional Import per Worker [0,.5] (.5,1] (1,1.5] (1.5,2] (2,2.5] (2.5,3] (3,3.5] (3.5,4] (4,6] [0,.5] (.5,1] (1,1.5] (1.5,2] (2,2.5] (2.5,3] (3,3.5] (3.5,4] (4,6] (a) 1997-2002 (b) 2002-2010 Notes: The figure shows a ten year equivalent change in import per worker for the period between 1997 and 2002 and 2002 and 2010. Figure 2 provides a geographical representation of IP W UK rt for 2002 and 2010 for all regions. The figure also clearly illustrates the extent of substantial geographical variation in the two time periods. On average, the change in regional import per worker was 960 pounds between 1997 and 2002, and it increased to 1,160 pounds for the period between 2002 and 2010, which is a 20.8 per cent increase. Among the most populous regions, the most exposed include Northamptonshire, South and West 9

Derbyshire, and Leicestershire CC and Rutland. 4 Finally, data on regional characteristics such as the proportion of female workers and a proportion of manufacturing workers is obtained from OECD regional statistics (OECD, 2017). 4. Econometric Methodology Since our objective is to investigate the causal relationship between trade shocks and wage inequality, we base our analysis on a quantile regression approach. Particularly, the quantile regression helps us to investigate the impact of trade shocks at different levels of the wage distribution. This is done by analysing the presence of significant changes at different quantiles of the wage distribution. It particularly help us to investigate whether import competition exacerbates or cushions wage inequality. Therefore, we adopt a region IV quantile approach developed by Chetverikov et al. (2016) to identify the effect of trade shocks on wage inequality. The econometric model is given by: Q wir v ir,x r,µ r (τ) = v irγ(τ) + x rβ(τ) + µ r (τ), τ (0, 1) (2) where Q(τ) represents τ th conditional quantile; w ir stands for dependent variable e.g. log weekly earnings of an individual i in region r; v ir represent individual level covariates that affect the dependent variable; γ(τ) is the τ th quantile coefficient estimates for individual covariates; x r corresponds to regional level covariates; β(τ) is a coefficient for region level covariates; and µ r (τ) represent region level unobservables. Given the general model, the identification of our variable of interest, β(τ), which is region level treatment effect, involves two steps. In the first step, we undertake quantile regressions using individual level outcome as dependent variable on individual characteristics for each region separately. This is given by: w ir (τ) = v irα r (τ) + u ir (τ) with Q τ (w ir v ir, x r, α r ) = v irα r (τ) (3) where, u ir (τ) is individual level iid error term. The coefficient estimate of the quantile 4 Table A1 of the appendix presents the top five most exposed regions out of the most fifty highest regions in terms of their working-age population for the year 2002 and 2010. The table also presents median, mean and standard deviation for the changes in import exposure. 10

regression, ˆα τ, solves the following equation for each region r. ˆα r (τ) argmin α A 1 n ni=1 [ρ τ (w ir v irα)] (4) where ρ τ (.) is known as the check function and can be rewritten as: τu τi if u τi 0 ρ τ (u τi ) = (τ 1)u τi if u τi < 0. (5) After solving equation (4) using quantile regression we get coefficient estimator for k individual level covariates, α r,k (τ), and the constant term, α r,1 (τ) = x rβ(τ) + µ r (τ), which corresponds to region fixed effect from equation (2). In the second step, we use the saved coefficient estimator from the constant term as a dependent variable and regress it on region level covariates, x r, to find estimator for our variable of interest, β(τ). We use a 2SLS estimation in the second step because, as we will discuss broadly below, our variable of interest is endogenous. Therefore, we regress α r,1 (τ) on x r using z r as an instrument. The corresponding empirical implementation of the above in our case follows three steps. In the first step we control for individual characteristics; in the second step we estimate the quantile; and in third step we estimate 2SLS to identify the variable of interest. To control for individual characteristics, we estimate Mincer-type wage equation for each region independently using the following specification. lnw irt = α 1 + α 2 age irt + α 3 age 2 irt + α 3 sex irt + α 4 occup irt + α 5 ft irt + α 6 manuf irt + ε irt, with E(ε irt v irt = 0) where the dependent variable lnw ir indicates the log of weekly earnings, log hourly wage or log total paid hours worked of an individual i in region r; individual level covariates, v ir, include age, age square, dummy for sex, which is one if the individual is male, nine occupational categories as a proxy for education, full time marker, and manufacturing marker; and ε ir represent standard regression residual. The above regression controls for observed individual characteristics. Following this, we prepare the dependent variable such that it represents regional changes in residual wage at each quantile levels. The change in residual wage at different quan- (6) 11

tiles will be between 2002-1997 and 2010-2002. This gives us regional level residual wage at each quantile, ε r (τ), for the years under consideration. And, thus change in residual wage between the years, ε r (τ). In the third step, we follow a 2SLS estimation approach and regress the change in region specific residual wages across quantiles, ε r (τ), on the change in import per worker, IP Wr UK, other region level covariates, x r 1, time dummy and region fixed effects. Here, we use change in the import of other counties, z r, as an instrumental variable, and identify β(τ). The empirical specification can be given by: ε r (τ) = IP W UK r β(τ) + x r γ(τ) + δ r (τ) + η t (τ) + µ r (τ), (7) where ε r (τ) indicates decadal equivalent changes in τ th quantile residual wage of region r ; and x r represent region level covariates other than a measure of change in import exposure. These covariates include the beginning of period regional characteristics such as percentage of employment in manufacturing, the percentage of employment among women, and the percentage of employment in routine occupations. 5 The last three terms, that is, δ r (τ), η t (τ) and µ r (τ), respectively, indicate NUTS 1 region fixed effects, time dummy for period 2002-2010 and the random iid error term. In the baseline empirical regressions, the dependent variables are weekly earnings, hourly wage and total paid hours worked. Unless specified, the main specifications are in changes; the standard errors are clustered at NUTS 2 level; regressions are weighted by their start of period population share, and all regressions include a constant term. Individuals are aggregated by region thus the number of observations in the regression tables reflect this aggregated figure. We have two periods thus the maximum sample size in our analysis is 256. Given that our main outcome of interest is wage distribution rather than an average wage, we combine the quantile regression with IV approach by which we can analyse wage inequality while correcting for the problem of endogeneity which arises from unobserved labour demand shocks. Studies that analyses labour market outcomes of trade shocks face potential endogeneity problems. This is because unobservable characteristics of the local labour market are likely to be correlated with the trade shocks. Specifically, the coefficient of the change in import exposure which 5 An index measuring routine task-intensity (RTI) of occupations for each region is calculated following Autor, Levy, and Murnane (2003) and Goos, Manning, and Salomons (2014)... 12

measures the impact of the change in exposure on the percentage change in wage is endogenous. Therefore, estimating equation (7) using OLS is problematic because an unobservable demand shock in the error term is likely to be correlated to changes in import exposure. To address the endogeneity problem, ADH introduce Bartik type instrument where they use change in imports from China in other developed countries as an instrument. They argue this external instrument is exogenous to the labour market outcomes but is correlated with trade shocks. Following ADH, we construct import exposure of seven developed countries, namely, Australia, Canada, Japan, Korea, New Zealand, Singapore, and the United States of America. 6 The instrument is constructed by making two changes to the import exposure measure of the UK as defined in equation (1). First, the regional share of manufacturing workers out of UK s national employment is lagged by six years 7 to avoid reverse causality, that is, current wage and employment may respond to expected trade exposure. Second, we use a change of import values from the six countries instead of a change of import from the UK. The equation below illustrates this fact: IP W OT H rt = j Emp rjt 6 Emp jt 6 IMP oth jt Emp rt 6, (8) where IP W OT H Emp rjt 6 Emp ujt 6 rt is a change in import exposure of other developed countries, is the six years lagged share of industry j employment in region r out of the UK s national employment of industry j, IMPjt oth is a change in the import of industry j by other countries, and Emp rt 6 is six year lagged level of employment in region r. There are also other measures of exposure which are used as alternatives in robustness checks 8. 6 The results are robust to include Norway and Switzerland or remove the USA from the group of other developed countries. Also, to avoid correlation in demand and supply shocks with the UK labour market, we excluded countries of the European Union. 7 To avoid reverse causality from anticipated import shock on current exposure, we use lagged employment share. Six years is the longest lag available in the data. 8 Other measures include net import and implied comparative advantage using residual of the gravity model. 13

5. Results 5.1. Basic Results Before presenting the main quantile results, it is indicative to consider an analysis at the average level. In fact, these results can be later compared with the findings of ADH. Moreover, we can also test the strength of our instrument and show the underlying relationship between openness and wage inequality at the average. Figure 3. Change in average wage response to change in import per worker Change in Log Weekly Wage -.5 0.5 1 0 2 4 6 Change in Trade Exposure Change_in_log_weekly_wage Fitted values In the above figure, we provide a scatter plot with a fitted line for a change in decadal equivalent mean log weekly earnings and change in regional imports per worker in the pooled data. Here regions are weighted by their start of decade population shears and the size of the bubble indicates their respective sizes. The slope of the fitted line is -0.039, indicating the inverse relationship between change in exposure and wage growth. 14

Moving to the econometric results, Table 2 provides 2SLS estimation results of the change in mean log weekly wage in a region on the change in import per worker. The six model specifications differ in terms of added controls. However, in all specification cases, the signs and significance of the estimated coefficient of our variable of interest remain the same. Table 2. Dependent Variable: Change in the log of average weekly earnings First Stage (1) (2) (3) (4) (5) (6) IPW OT r H 0.119 0.075 0.083 0.075 0.071 0.071 (0.008) (0.009) (0.007) (0.01) (0.006) (0.006) Time dummy -0.428-0.371-0.221-0.222 (0.054) (0.052) (0.077) (.077) L_female_share -0.229-0.141 (1.029) (1.074) L_routine_share 2.308 2.267 (0.711) (0.717) L_manuf_share 1.258 (2.857) Region FE No No No Yes Yes Yes F 235.9 73.34 137.5 208.2 122.9 120.2 Partial R 2 0.622 0.606 0.588 0.507 0.438 0.436 Second Stage IPW UK r -1.569-1.543-1.549-1.265-1.714-1.818 (2.760) (1.906) (0.977) (1.540) (2.893) (2.912) Time dummy -26.996-27.059-28.213-28.363 (2.666) (2.743) (3.958) (3.944) L_female_share -66.255-58.483 (83.668) (81.930) L_routine_share -25.364-28.742 (24.338) (24.345) L_manuf_share 110.82 (114.305) N 128 128 256 256 256 256 R 2 0.002. 0.536 0.546 0.552 0.553 Notes: All the regressions include constants and NUTS2 level clustered standard errors. * Denotes significance at the 10% level. ** Denotes significance at the 5% level. *** Denotes significance at the 1% level. 15

The first panel shows the first stage regression results. As indicated by its statistical significance, our instrument, changes in import per worker of the other developing countries, is highly correlated with changes in import exposure of the UK. The instrument also explains a significant amount of variation in the endogenous variable as indicated by a relatively high partial R 2. Also under all model specifications, the F-tests are above critical value ensuring the absence of a weak instrumental variable problem. The lower panel of Table 2 shows that the effect of a change in import exposure on mean log wage is negative but insignificant. Column one and two present regression results for the two periods separately. The estimated coefficients tell us a 1000 increase in import per worker from China during a decade reduces average wage by 1.57 log points, however, the estimates are not statistically different from zero. Column three presents regression results where we include a time dummy for the period 2002-2010. Here, the time dummy is negative and significant. The coefficient indicates that there is a 27 log points reduction in the decadal equivalence relative growth of average wage in the second time period relative to the first period. In column four, we add regional fixed effects along with a time dummy. In the specification of the fifth column, we add the beginning of period regional characteristics: lagged share of female workers and lagged share of employment in routine jobs. Both variables seem to have no significant effect on change in log average wage. And finally, in column six, we add lagged shares of manufacturing employment as an additional control. Similarly lagged share of manufacturing wage has no significant effect on changes in log average wage. The results at the average are in line with studies by Balsvik, Jensen, and Salvanes (2015) and Edward and Lawrence (2010), who also find no significant effect of import competition on mean wages. Although a simple scatter plot in Figure 1 and the regression results at the average help us to check for validity of our instrument and observe the general pattern, they are inadequate in telling the whole story. This is because the causal link between the change in wage and change in import per worker can be different at different parts of the wage distribution. Moreover, we need to base our analysis on methods such as quantile regression to answer our main research question: the effect of import competition on wage inequality at the level of local labour market. 16

Table 3. Models before controlling for individual characteristics Dependent Variable: Change in log of weekly earning (1) (2) (3) (4) (5) (6) (7) (8) (9) 10 th 20 th 30 th 40 th 50 th 60 th 70 th 80 th 90 th IPW UK r -2.249-3.269-1.726-1.644-1.148 0.442 0.0205 0.174 1.041 (2.666) (2.570) (1.912) (1.490) (1.926) (1.703) (1.303) (1.560) (1.581) L_manuf_share 222.0 207.2 123.7 121.1 120.2 136.2 144.8 123.3 210.0 (218.677) (158.928) (122.898) (106.127) (112.751) (100.729) (99.447) (92.018) (111.788) L_female_share 36.48 7.300 11.89 12.90 50.03 44.44 10.84-9.946 1.610 (60.451) (53.042) (45.418) (34.419) (35.251) (32.716) (26.970) (26.291) (36.257) L_routine_share -103.5-60.73-37.82-4.920 2.840-5.691-9.599-19.44-18.13 (45.669) (40.286) (32.665) (26.377) (26.126) (20.088) (17.155) (14.294) (13.095) Time dummy -43.17-33.25-24.78-21.50-20.83-20.82-21.35-22.47-24.49 (5.308) (4.281) (2.796) (2.695) (2.572) (2.001) (1.724) (1.552) (1.624) Region FE yes yes yes yes yes yes yes yes yes F 125.0 125.0 125.0 125.0 125.0 125.0 125.0 125.0 125.0 Partial R 2 0.373 0.373 0.373 0.373 0.373 0.373 0.373 0.373 0.373 R 2 0.354 0.352 0.400 0.442 0.500 0.483 0.554 0.593 0.571 N 256 256 256 256 256 256 256 256 256 Notes: All specifications include constants and NUTS2 level clustered standard errors. For all quantile regression, we control for start of period region characteristics such as share of manufacturing employment, the share of female in workers, the share of employment in routine workes, region fixed effects and a time dummy for the period 2002-2010. * Denotes significance at the 10% level. ** Denotes significance at the 5% level. *** Denotes significance at the 1% level. In table 3, we present regression result from group quantile IV without controlling for individual characteristics. In all of the analysis below, we follow similar model specifications as the last column of Table 2. As we can see from the table, although there is a change in coefficient estimate from negative to positive, as we go from the lower to the higher quantiles of the wage distribution, we do not find evidence to argue that a change in import exposure contributes to the rise in wage inequality. The first panel of Figure 4 shows the plot of estimated coefficients of the change in regional import per worker at different quantiles (by 5 percentile increment) with respective 90 and 95% confidence intervals. These regression specifications closely followed the study by Chetverikov et al. (2016) for the US. While they find evidence for the causal effect of an increase in import competition on wage inequality; from this specification, we do not have enough statistical power in our coefficient estimates to argue the presence of a causal link between wage inequality and a change in import exposure in the UK. However, since individual characteristics are important determinants of wage, we exploit the richness of ASHE data and control for individual characteristics in the wage equation. And once we control for individual characteristics, we find evidence for a causal effect of import competition on the rise in wage inequality. Table 4 (also panel b of Figure 2) presents the regression result of change in import exposure on 17

Change in log of weekly wage -10-5 0 5 Average Estimate Estimate with individual controls without controls and on average Figure 2. Estimation coefficients and confidence intervals for regression on average and quantile with and without controls for individual characteristics. Dependent variable is log of weekly wage and estimation is on all workers. Change in log of weekly wage -6-4 -2 0 2 Estimate Table 4. Models after controlling for individual characteristics Dependent Variable: Change in log of weekly earning (1) (2) (3) (4) (5) (6) (7) (8) (9) 10 th 20 th 30 th 40 th 50 th 60 th 70 th 80 th 90 th IPW UK r -0.202-0.891-0.745-0.387 0.139 0.305 0.204-0.109 0.440 (0.727) (0.321) (0.377) (0.322) (0.247) (0.345) (0.380) (0.531) (0.740) L_manuf_share -53.18-78.82-43.11-26.49-43.66-34.90-5.607-3.854-41.62 (41.982) (26.544) (26.927) (21.150) (25.940) (22.920) (22.277) (39.004) (89.297) L_female_share 0.915-12.13-0.948 3.039 5.004 13.67 3.470-19.62-9.033 (17.338) (11.388) (8.032) (6.729) (6.769) (6.874) (8.434) (9.040) (15.582) L_routine_share -0.0560 9.879 4.139 4.177-1.207 2.343-2.178-2.899 5.231 (14.440) (5.867) (4.768) (4.508) (5.522) (5.335) (6.106) (8.049) (12.098) Time dummy -4.953-1.810-0.574 0.160 0.193 0.834 1.237 2.795 7.008 (1.166) (0.571) (0.538) (0.445) (0.469) (0.561) (0.632) (0.737) (0.957) Region FE yes yes yes yes yes yes yes yes yes F 120.2 120.2 120.2 120.2 120.2 120.2 120.2 120.2 120.2 Partial R 2 0.436 0.436 0.436 0.436 0.436 0.436 0.436 0.436 0.436 R 2 0.195 0.159 0.062 0.045 0.037 0.066 0.105 0.163 0.265 N 256 256 256 256 256 256 256 256 256 Notes: All specifications include constants and NUTS2 level clustered standard errors. For all quantile regression, we control for start of period region characteristics such as the share of manufacturing employment, the share of female in workers, the share of employment in routine workes, region fixed effects and a time dummy for the period 2002-2010. * Denotes significance at the 10% level. ** Denotes significance at the 5% level. *** Denotes significance at the 1% level. changes in weekly wages after we control for individual characteristics such as age, sex, occupation. We find a negative effect of change in import exposure on change in weekly wage for lower quantiles, while the precision of our estimates at the bottom is not strong. But, the effect is significant for those between the 20 th and 35 th percentile of the wage distribution. For instance, for those at the 20 th percentile of the 18

wage distribution, a 1000 pound increase in regional import per worker is estimated to reduce their weekly wage by 0.89 log points. The effect on the higher quantiles becomes positive, with the exception for those at the 80 th percentile, albeit they are not significantly different from zero. This is in line with the findings by Autor et al. (2014) who find bigger effect at the lower part of the wage distribution for the US. To identify the underlaying mechanism of the wage effect, we disentangle the wage effect into its price and quantity components, that is, hourly wage and total hours worked. This is important to investigate the transition mechanisms through which import competition causally affects wage inequality. In Table 5, we present the results for the change in log of real average hourly earnings on different covariates. We include the same regional covariates as the above regression. The table shows that change in import per worker has both positive and negative effect for the different part of the wage distribution. Particularly, it is negative at the 30 th, 40 th, 80 th and 90 th while insignificant for all at conventional levels. At the other quantile of the wage distribution, it is positive with a 10% significance for those at the 60 th and 70 th quantiles. Table 5. Models after controlling for individual characteristics Dependent Variable: Change in log of hourly wage (1) (2) (3) (4) (5) (6) (7) (8) (9) 10 th 20 th 30 th 40 th 50 th 60 th 70 th 80 th 90 th IPW UK r 0.218 0.001-0.199-0.089 0.054 0.386 0.457-0.052-1.491 (0.777) (0.456) (0.323) (0.208) (0.213) (0.208) (0.275) (0.525) (0.775) L_manuf_share -36.64-33.57-17.27-13.24 9.246 23.75 26.89 17.00 28.52 (42.271) (36.983) (24.156) (14.288) (14.880) (18.050) (19.999) (27.171) (35.471) L_female_share 23.39 8.437 4.831 11.29 8.060 3.736-4.863-23.46-51.19 (19.077) (13.734) (9.166) (6.288) (4.216) (4.500) (8.065) (12.853) (19.821) L_routine_share 11.03 1.945-0.906 0.206-6.060-9.879-6.197 1.092 11.09 (8.791) (10.181) (8.911) (6.319) (5.103) (4.049) (5.202) (6.423) (11.586) Time dummy 0.082-0.910-0.850-1.274-1.404-1.947-2.484-1.897 0.141 (1.071) (0.737) (0.504) (0.425) (0.375) (0.278) (0.436) (0.750) (0.986) Region FE yes yes yes yes yes yes yes yes yes F 120.2 120.2 120.2 120.2 120.2 120.2 120.2 120.2 120.2 Partial R 2 0.436 0.436 0.436 0.436 0.436 0.436 0.436 0.436 0.436 R 2 0.059 0.067 0.051 0.109 0.099 0.152 0.200 0.142 0.081 N 256 256 256 256 256 256 256 256 256 Notes: All specifications include constants and NUTS2 level clustered standard errors. For all quantile regression, we control for start of period region characteristics such as share of manufacturing employment, the share of female in workers, the share of employment in routine workes, region fixed effects and a time dummy for the period 2002-2010. * Denotes significance at the 10% level. ** Denotes significance at the 5% level. *** Denotes significance at the 1% level. Moreover, Table 6 presents the quantile IV regression results for the change in log of average total paid hours worked on changes in import per worker. The results for average total paid hours worked is negative for those at the lowest and highest 19

Change in log of hourly wage -4-2 0 2 4 Change in log of total hours worked -6-4 -2 0 2 4 Estimate Estimate with individual controls with individual controls Figure 3. Estimation coefficients and confidence interval for regression on average and quantile with and without controls for individual characteristics. Dependent variable is log of weekly wage and estimation is on all workers. part of the distribution of hours worked but it is not significantly different from zero. On the other hand its effect on those who are at the upper middle quantiles declines but it is still insignificant. From these results, we cannot argue that the adjustment is through a reduction in total hours worked or a reduction in hourly earnings. Table 6. Models after controlling for individual characteristics Dependent Variable: Change in log of total hours worked (1) (2) (3) (4) (5) (6) (7) (8) (9) 10 th 20 th 30 th 40 th 50 th 60 th 70 th 80 th 90 th IPW UK r -0.358-0.123-0.069-0.217 0.002 0.139 0.191 0.223-0.256 (0.812) (0.434) (0.358) (0.219) (0.216) (0.331) (0.436) (0.596) (0.850) L_manuf_share -72.68-17.05-2.977-6.104-17.68-23.78-16.53 12.48 28.11 (32.860) (15.665) (15.742) (12.982) (11.163) (12.563) (19.419) (30.610) (50.698) L_female_share -39.46-18.43-6.899-2.278 2.042 9.825 17.24 32.05 18.74 (18.847) (9.320) (7.482) (4.457) (2.799) (7.374) (10.701) (14.445) (22.637) L_routine_share -6.198 0.903 2.008 6.349 1.944 0.223-4.051-2.313 17.03 (9.475) (3.707) (3.412) (3.417) (3.493) (3.472) (4.634) (6.660) (13.241) Time dummy -5.997-1.490-0.107 0.402 0.682 0.797 0.662 2.359 7.097 (1.346) (0.595) (0.505) (0.318) (0.299) (0.463) (0.567) (0.717) (1.279) Region FE yes yes yes yes yes yes yes yes yes F 120.2 120.2 120.2 120.2 120.2 120.2 120.2 120.2 120.2 Partial R 2 0.436 0.436 0.436 0.436 0.436 0.436 0.436 0.436 0.436 R 2 0.347 0.168 0.044 0.030 0.081 0.089 0.086 0.153 0.264 N 256 256 256 256 256 256 256 256 256 Notes: All specifications include constants and NUTS2 level clustered standard errors. For all quantile regression, we control for start of period region characteristics such as the share of manufacturing employment, the share of female in workers, the share of employment in routine workes, region fixed effects and a time dummy for the period 2002-2010. * Denotes significance at the 10% level. ** Denotes significance at the 5% level. *** Denotes significance at the 1% level. 20

5.2. Regression on Sub-samples In this section, we present regression results on different subsample of workers, namely, male, female, manufacturing, non-manufacturing, full time and part time workers. Here, all the results are after we control for individual characteristics. As panel A of Table 7 shows there is no significant effect of change in import per worker on change in wage of female workers throughout the wage distribution. However, as shown in panel B, for male workers, there is a negative and significant effect for those between the 15 th and 45 th percentiles of the wage distribution. In both cases, we control for individual characteristics other than gender. Table 7. Models after controlling for individual characteristics on the subgroups Dependent Variable: Change in log of weekly wage (1) (2) (3) (4) (5) (6) (7) (8) (9) 10 th 20 th 30 th 40 th 50 th 60 th 70 th 80 th 90 th Panel A. For Female Workers IPW UK r 0.195-0.597-0.529-0.325 0.249 0.365 0.255 0.787 0.232 (1.281) (0.725) (0.464) (0.467) (0.385) (0.551) (0.607) (0.966) (1.274) R 2 0.144 0.094 0.108 0.059 0.037 0.083 0.092 0.149 0.166 Panel B. For Male Workers IPW UK r 0.095-1.245-1.062-0.372-0.109-0.000-0.855-0.530 0.351 (0.539) (0.508) (0.361) (0.389) (0.437) (0.464) (0.457) (0.520) (0.961) R 2 0.126 0.073 0.033 0.033 0.041 0.036 0.071 0.133 0.164 Panel C. For Manufacturing Workers IPW UK r -2.659-1.754-0.953 0.467 1.176 1.786 1.942 0.417 2.669 (1.287) (1.209) (0.795) (0.609) (0.493) (0.668) (0.870) (0.987) (1.407) R 2 0.054 0.046 0.044 0.020 0.027 0.049 0.066 0.054 0.053 Panel D. For Non-manufacturing Workers IPW UK r -0.517-0.655-0.665-0.321 0.0291-0.162-0.152 0.122 0.0499 (0.818) (0.429) (0.458) (0.345) (0.311) (0.420) (0.455) (0.680) (0.974) R 2 0.162 0.104 0.047 0.042 0.044 0.050 0.070 0.157 0.198 Panel E. For Full-time Workers IPW UK r -0.820-0.836 0.0714-0.0508 0.195 0.434 0.677 0.140-0.119 (0.460) (0.414) (0.421) (0.238) (0.253) (0.275) (0.338) (0.467) (0.656) R 2 0.052 0.054 0.025 0.049 0.062 0.114 0.205 0.136 0.104 Panel F. For Part-time Workers IPW UK r -2.149-2.111-0.537-0.756-0.437-0.002-0.764-0.647 0.157 (2.359) (1.465) (1.074) (0.783) (0.826) (1.059) (1.185) (1.362) (1.788) R 2 0.173 0.277 0.089 0.045 0.209 0.234 0.299 0.262 0.130 Notes: All columns report second stage estimated coefficent for the change in import per worker from 2SLS estimation. All the regression include constants and NUTS2 level clustered standard errors. For all quantile regression, we control for start of period region characteristics such as share of manufacturing employment, share of female in workers, share of employment in routine workes, region fixed effects and a time dummy for the period 2002-2010. * Denotes significance at the 10% level. ** Denotes significance at the 5% level. *** Denotes significance at the 1% level. Panel C and D shows the regression result for manufacturing and nonmanufacturing workers. For Manufacturing workers, we find negative and significant effects for those at the 10 th quantile. While, for those at the 50 th, 60 th and 70 th percentile, the effect is positive and significantly different from zero at 5% level. We 21