Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Essays on trade liberalization and labour market outcomes Townsend, James Herbert 2002

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


831-ubc_2002-750949.pdf [ 6.88MB ]
JSON: 831-1.0090629.json
JSON-LD: 831-1.0090629-ld.json
RDF/XML (Pretty): 831-1.0090629-rdf.xml
RDF/JSON: 831-1.0090629-rdf.json
Turtle: 831-1.0090629-turtle.txt
N-Triples: 831-1.0090629-rdf-ntriples.txt
Original Record: 831-1.0090629-source.json
Full Text

Full Text

ESSAYS ON T R A D E L I B E R A L I Z A T I O N A N D L A B O U R M A R K E T O U T C O M E S by J A M E S H E R B E R T T O W N S E N D B.Sc.(hon.), The University of Calgary, 1994 M.A. , The University of Calgary, 1995 A THESIS S U B M I T T E D IN P A R T I A L F U L F I L M E N T OF T H E R E Q U I R E M E N T S F O R T H E D E G R E E OF D O C T O R OF P H I L O S O P H Y in T H E F A C U L T Y OF G R A D U A T E STUDIES (Department of Economics) We accept this thesis as conforming to the required standard T H E U N I V E R S I T Y OF BRITISH C O L U M B I A July 2002 ©James Herbert Townsend, 2002 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of \z. C o i a o ^ i The University of British Columbia Vancouver, Canada Date 9 ^ 2£>C2-DE-6 (2/88) A b s t r a c t This thesis uses a comprehensive data set to examine the relationship between Canadian labour market outcomes and several changes in the policy environment. The data set, spanning the period 1981-98, is compiled from a number of compara-ble surveys and contains information on the demographics and job characteristics of individual workers. The first chapter examines the impact that the tariff reductions of the Canada-U.S. Free Trade Agreement (CUSFTA) had on the inter-industry wage structure in the goods producing sector. Previous studies use industry-level data and con-sequently are unable to control for either differences in worker composition or divergent wage trends for different worker types. These studies find that tariff cuts either had no effect or increased the relative wages of workers in impacted sectors. In contrast, I use data with information on worker characteristics and find that the relative wages of non-union workers in impacted industries decreased. The second chapter investigates the link between the CUSFTA tariff reductions and several labour market outcomes that are potentially linked to industrial pro-ductivity. In particular, I examine whether tariff reductions are related to changes in the (i) the size of firm a worker is likely to be employed with, (ii) the proba-bility that a worker will be represented by a union, and (iii) the mean skill level of workers. Although I find evidence that these outcomes have changed over time, none of them seem to be linked directly to CUSFTA. The final chapter, co-authored with David Green, examines the extent to which the declining market outcomes of successive cohorts of job entrants in Canada can be accounted for by changes in the minimum wage, unionization rate, and industrial composition of employment. A flexible density estimator is used, which allows for a comparison between cohorts across the entire wage distribution. The main findings are that for males, changes in unionization and industrial composition can account for about a quarter of the decline in wage outcomes for new job entrants between 1998 and 1981. Similar results are found for females; in addition, the minimum wage provides a "wall" against further erosion for more recent cohorts of entrants. ii Table of Contents Abstract 1 1 Table of Contents i i i List of Tables v i List of Figures v i i i 1 Overview and Summary 1 2 Free Trade and The Canadian Wage Structure: A Mic roda ta Analysis 4 2.1 Introduction 4 2.2 Trade Liberalization and Wages: Theory and Empirical Evidence . 6 2.2.1 Rent-sharing 7 2.2.2 Compensating differentials 9 2.2.3 Industry-specific human capital 9 2.2.4 Adjustment differentials 10 2.2.5 Summary of Predictions 10 2.3 Empirical Framework 11 2.3.1 Existing Empirical Studies 13 2.4 Data 14 2.5 Empirical results 16 2.5.1 Wage and tariffs over the period 1981-1998 17 2.5.2 Workforce Characteristics by Industry 18 2.5.3 Regression results 20 2.5.4 Past and Present Results: A Reconciliation 24 2.6 Interpreting the results 28 2.6.1 Collective bargaining and rent-sharing 28 2.6.2 Specific human capital 29 2.7 Summary and Conclusions 30 3 Productivi ty, Labour Market Outcomes and Trade Liberalization: A Microda ta Analysis 42 3.1 Introduction '. 42 iii 3.2 Theory 44 3.2.1 Scale 44 3.2.2 Unionization 46 3.2.3 Skills Content 48 3.3 Data 49 3.4 Methodology 50 3.5 CUSFTA and Selected Outcomes 52 3.5.1 Plant Scale 52 3.5.2 Unionization 54 3.5.3 Skills allocation 55 3.6 Discussion and Conclusion 58 4 The Sources of Declining Entry Wages for the Less Educated in Canada 65 4.1 Introduction 65 4.2 Previous Literature 68 4.2.1 An Empirical Framework 68 4.2.2 Previous findings 70 4.3 Data 75 4.4 Basic Patterns in the Canadian Data 77 4.4.1 Mean Wages 77 4.4.2 Covariates 85 4.4.3 Estimation: Mean Regression Models 87 4.5 Estimation of Conditional CDFs 89 4.5.1 An Estimator for CDFs Conditional on Covariates 89 4.6 Results 93 4.6.1 Estimated Marginal Covariate Impacts 93 4.6.2 Decompositions of Shifts in the CDF 96 4.7 Conclusions 101 A Auxiliary regressions 121 Bibliography 124 iv List of Tables 2.1 Average wages and tariffs, no individual controls 35 2.2 Worker demographics, "tradables" sector only 36 2.3 Union coverage and gender breakdown by industry, selected years . 37 2.4 Proportion of unskilled and young workers by industry, selected years 38 2.5 Wage impact of CUSFTA, less skilled, non-union sector 39 2.6 Wage impact of Canadian Tariff, by skill and union status 39 2.7 Wage impact of CUSFTA, by skill and union status 40 2.8 The Effects of Controlling for Worker Characteristics 40 2.9 Contribution of Changes to Composition and Returns of Worker Characteristics 41 3.1 Employment share by firm size, selected industries and years . . . . 60 3.2 Probit analysis of likelikhood of being employed by a large firm . . 61 3.3 Employment share by firm size and tariffs 62 3.4 Effect of tariffs on union status 62 3.5 Skills distribution by industry 63 3.6 Effect of tariffs on skills distribution 63 3.7 Simulated skills distribution, high tariff industries, 1998 64 3.8 Simulated effect of tariffs on skill distribution, high tariff industries, 1988 64 4.1 Differences in Log Wages By Cohort and Year: Measured Relative to 25-34 Year Olds in 1981 116 4.2 Mean Log Wage Regressions, 1981, 1989 and 1998 Data 117 4.3 Covariate Distributions for 1981, 1989 and 1998 Job Entrant Cohortsll8 4.4 Job Entry Cohort Effects from Mean Log Wage Regressions High School Males 119 v 4.5 Job Entry Cohort Effects from Mean Log Wage Regressions High School Females 120 A . l Mean Log Wage Regressions: Males Additional Specifications Using Data From A l l Years 122 A.2 Mean Log Wage Regressions: Males Additional Specifications Using Data From A l l Years 123 vi List of Figures 2.1 Average Tariff Rates, 1981-1998 33 2.2 The 1988 Tariff Structure 33 2.3 The Gender Differential 34 3.1 Union Density, by Industry Groupings 60 4.1 Mean Log Wages by Age Group, Year and Job Tenure, High School Men 102 4.2 Mean Log Wages by Age Group, Year and Job Tenure, High School Women 103 4.3 Mean Log Wages by Job Entry Cohort, High School Men 104 4.4 Mean Log Wages by Job Entry Cohort, High School Women . . . . 105 4.5 Effects of minimum wage on the wage distribution, High school men, 25-34 years old, manufacturing 106 4.6 Effects of union coverage on the wage distribution, High school men, 25-34 years old, manufacturing 107 4.7 Effects of industry coverage rate on the wage distribution, High school men, 25-34 years old, manufacturing 108 4.8 Effects of industry on the wage distribution, High school men, 25-34 years old 109 4.9 Effects of entry cohort on the wage distribution, High school men, 25-34 years old, manufacturing 110 4.10 Effects of entry cohort on the wage distribution, no minimum wage, High school men, 25-34 years old, manufacturing I l l 4.11 Effects of entry cohort on the wage distribution, , High school women, 25-34 years old, manufacturing 112 4.12 Difference between years, actual and fitted, high school men . . . . 113 vii 4.13 Decomposition of Difference between Wage Distributions, High School Men, Ages 25-34, 1981 and 1998 114 4.14 Decomposition of Difference between Wage Distributions, High School Women, Ages 25-34, 1981 and 1998 115 viii Chapter 1 Overview and Summary There have been a number of changes in the economic environment in the last twenty years that have potentially important implications for labour market out-comes in Canada. Relevant developments include Canada's participation in various trade agreements, such as the Canada-U.S. Free Trade Agreement (CUSFTA) and the North American Free Trade Agreement (NAFTA), changes to minimum wage legislation, and revisions to the legislation governing unionization. In addition, there have been a variety of shifts in employment patterns, particularly for new job entrants. New jobs are less likely to be covered by a collective bargaining agreement. In additional, employment opportunities have shifted away from tra-ditional sectors like the primary goods sector to services. A l l these developments have potential implications for the wage distribution. This paper contributes to our understanding of these developments. In the first essay, I investigate the effects of the Canada-U.S. Free Trade Agree-ment (CUSFTA) on the wages of workers in the impacted industries. Previous work by Gaston and Trefler (1997) and Beaulieu (2000) found no connection be-tween the average weekly industrial earnings and the mandated tariff reductions. Unlike previous studies, I examine this issue using a detailed labour data set that I constructed from numerous micro-data sets. These data sets contain information on worker and job characteristics that were not available to earlier researchers. The inclusion of these extra variables allows me to control for compositional shifts within industries and divergent wage trends across diverse groups of workers. I proceed by estimating wage equations similar to those used by the previous researchers using only industry level data and find similar results. I gradually introduce extra information on worker characteristics and union status. By con-trolling for worker characteristics and allowing for different secular trends across 1 I indentifiable worker groups, I come to a different conclusion that that of the exist-ing literature. In particular, I find that tariff reduction had a detrimental impact on the wages of workers in formerly protected industries. For example, for men with a high school education or less, a one percentage point reduction in the appli-cable tariff resulted in a 0.7% reduction in the wages. In the most highly protected industries, this implies wage losses in excess of 10% over the period 1988-98. A l -though the wage effects were large, I find that only a small fraction of the workforce were employed by these industries. This suggests that CUSFTA cannot account for much of the increase in wage inequality observed in Canada over the last twenty years. In the second essay, I examine the role of CUSFTA in influencing a number of labour market outcomes that might potentially account for the productivity gains observed in the Canadian manufacturing sector over the period in which the CUSFTA-mandated tariff reductions were phased in. As Head and Ries (1999) document, although the number of establishments in manufacturing decreased in the years following the signing of CUSFTA, the output per firm increased. This observation is consistent with the notion that increased competition would result in larger firms which exploited increasing returns to scale. Ries and Head (1999) find little evidence the tariff reductions were directly responsible for the observed productivity gains. Trefier (2001) finds a correlation between the tariff reductions and observed increases in labour productivity. However, he finds no indications that scale effects are responsible, instead concluding that favourable firm turn-over explains most of the gains. As a complement to this literature, I examine several labour market outcomes that are potentially related to productivity. In particular, I look at within indus-try developments in employment shares across firms of varying sizes, unionization, and educational attainment of the workforce. A shift in employment from small to large firms would be consistent with exploitation of scale economies. The lit-erature on unions has suggested a number of implications for firm productivity, both positive and negative. Output per worker may increase with the average level of educational attainment. My findings are mostly negative. Although there was a shift from small to large firms in the industries that were formerly highly protective, there is no statistically significant connection to the tariff reductions. Unionization rates decrease and educational attainment increases in manufactur-ing in the 1990's. Although these trends may contribute to the observed gains 2 in productivity in manufacturing, there is no evidence that these developments occurred any differently in the industries that faced the highest tariff cuts. In the third essay, co-authored with David Green, we examine developments in employment patterns, unionization, and the minimum wage in an attempt to fur-ther understand why labour market outcomes have been declining for subsequent cohorts of job entrants. This work builds on work by Beaudry and Green (2000), in which the authors discover that the wage-experience profiles have been shifting down for subsequent birth cohorts in Canada. The wage gap between younger and older men has widened considerably in Canada over the last twenty years. Some have thought this to be evidence of an increase in the premium to experience. As age and experience are closely correlated, this would show up as a widening in the gap between the earning of the young and old. It should also result in the lifetime wage profile of individuals becoming steeper over time. Beaudry and Green (2000) find no evidence that the wage profile has become steeper over time. A comparison of age-earnings profiles across cohorts reveals instead that age-profiles are similar across birth cohorts but have been decreasing in levels with each successive cohort. An acceleration in the rate of decline among men with a high school education or less has resulted in the expansion of the gap between the young and old. In the third essay, we reorganize the cohort pattern according to cohorts of job entrants. We show that this alternative representation captures the wage patterns well. We then investigate how the labour market experiences of successive job entry cohorts has changed, with regards to industrial employment patterns, union representation, and the applicable level of minimum wage. We examine the extent to which different wage experiences across cohorts can be accounted for by these developments in the labour market. A flexible density estimator is used to estimate the effects of unionization, the minimum wage, and job structure on the wage distributions of various job entry cohorts. This methodology allows us to correctly account for the effects of union representation and the minimum wage, both of which have different implications over different parts of the wage distribution. Our findings are that cross-cohort differences in unionization and job structure can account for about half of the mean wage loss between the 1980 entry cohort and the 1998 entry cohort. The minimum wage mitigated the decline for females at the lower end of the wage distribution. 3 Chapter 2 Free Trade and The Canadian Wage Structure: A Microdata Analysis 2.1 Introduction The Canada-U.S. Free Trade Agreement (CUSFTA) was a major issue during the Canadian election campaign in 1988. Supporters argued that the agreement would bring about substantial welfare gains. A variety of computational gen-eral equilibrium models predicted that large gains in GDP would result from the deal.1 Oligopolistic domestic firms were expected to face increased competition from American firms, which would force down price mark-ups. Lower prices would force some firms out of the market, while the rest would respond to lower prices by increasing output to exploit scale economies. Increased access to the large Ameri-can market would allow Canadian firms to further realize economies of scale. These productivity gains would be accompanied by wage gains. In contrast, opponents argued that the Free Trade agreement would result in the the de-industrialization of Canadian. Jobs would be lost and wages would fall as firms relocated produc-tion to take advantage of American amenities such as lower taxes and labour laws which made union organization more difficult. Canada would specialize in the production of natural resources to supply to the American industrial machine. Thirteen years after the implementation of the Free Trade agreement and seven years after the implementation of the follow-up North American Free Trade Agree-ment (NAFTA), the debate over the consequences of these agreements is still 'See for example Cox and Harris (1984). 4 heated. There is little evidence that Canada has reaped the substantial gains that the computable general equilibrium models predicted. Gaston and Trefler (1997) found that the CUSFTA tariff reductions could account for about 15% of the unemployment experienced over the recessionary period from 1989-93. They found no evidence that the tariff reductions had any affect on wage growth over the same period. Beaulieu (2000) found that most of the employment losses were among less-skilled production workers, with white-collar workers in the impacted sectors being relatively unaffected. Trefler (2001) finds that the tariff cuts caused short-run job losses but resulted in increased labour productivity and associated wage gains for workers in the impacted sectors. Trefler argues that productivity increased as a result of favorable firm turn-over rather than as a result of scale economies. While academic studies have suggests relatively small losses and gains associated with CUSFTA, Bruce Campbell of the left-wing Canadian Centre for Policy Alternatives argues that the free trade agreements and associated policies were key contributors to the high employment and wage stagnation that were observed throughout- the early nineties.2 To further extend our knowledge of the labour market implications of trade liberalization, I re-examine the effects of the CUSFTA tariff reductions on the in-dustrial wage structure. I first review different models of trade and the predictions that they make about the impact of tariff reduction on trade liberalization. I then discuss the general empirical framework used to examine the relationship between wages and tariffs. In the process, I consider the extent to which this framework can be related back to trade theory. I then proceed to estimate a relationship between tariff reduction and wages, using an empirical framework that is similar to the one used by Gaston and Trefler (1997), and Beaulieu (2000). Unlike previous studies, I use micro-data which allows me to control for worker characteristics, along with other important determinants of wages such as union status. Unlike previous studies, I find that the tariff reduction had a substantial impact on the industrial wage structure for certain types of workers. For less-educated, non-union workers in the most impacted industries, free trade had a devastating 2See Campbell et al. (1999). Some commentators believe that the high employment of the early nineties resulted from the anti-inflationary interest rate policy that was implemented by the Bank of Canada. Campbell argues that this policy was pursued as part of a policy to "discipline the market" so that the Canadian labour market could compete with the American market. He argues that this adjustment was necessitated by CUSFTA. In his assessment, treating high interest rates as a phenomena that was completely seperate from C U S F T A results in under-stating the cost of liberalization. 5 effect.3 These workers received among the lowest wages within the goods-producing sector prior to the CUSFTA tariff reductions. Relative wages in these sectors fell even further behind as a result of the tariff reductions, with a one percentage point drop in the statutory rate against American imports resulting in a 0.7 % drop in the real wage in the impacted industry. In 1988, several industries had tariffs between 12.8% and 17.2%. My point estimate indicates that for these workers CUSFTA resulted in a decline in relative real wages of between 8% and 12%. Despite a strong direct effect on the industrial wage structure, the overall impact was small, as the 1988 tariff rate in the majority of industries was small. I argue that these results are consistent with a model in which liberalization affects the return to industry-specific human capital. This effect appears to be mitigated by union representation. My ability to control for the gender of workers appears to be important in accounting for why my results differ from those of the previous studies. Sectors that traditionally received protection employ a higher proportion of women. I find evidence that the male-female wage gap in the goods producing sector closed in the post-CUSFTA period; the combination of the closing of this gap, along with differences in employment shares across industries generated movements in the mean wages of industries that countered those movements generated by reducing tariffs. 2.2 Trade Liberalization and Wages: Theory and Empirical Evidence Predictions about the welfare implications of CUSFTA were largely based on computational general equilibrium models. These models, based on the "new" trade literature, emphasize imperfect competition and increasing returns to scale.4 In these models, tariff protection permits oligopolistic firms to raise prices. Higher prices encourage additional entry; new entrants "crowd" out the incumbents, re-sulting in reduced output per firm. In addition to the allocative inefficiency result-ing from oligopolistic mark-ups, the reduction in output at the firm level results in additional productive inefficiencies. In particular, with increasing returns to scale, the reduction in output increases the per unit cost of production. Under 3 As less skilled workers accounts for around 50% of employment in the sectors of interest over the study period, I emphasize my findings for this group. As will be seen, other groups were adversely impacted by the tariff reductions as well. 4See Hazledine (1990) for a review of these models. 6 these assumptions, reducing the tariff against imports will push down the price level. As the price level falls below average cost, some firms will exit, while the remainder will compensate for the lower price level by expanding production and thus reducing the average cost of production. CUSFTA would permit further pro-ductivity gains as Canadian firms gained access to the large American market and experienced further scale returns. Simulations of CUSFTA predicted substantial reallocation of labour across in-dustries, with employment expanding in some industries and contracting in others. In contrast, wage implications were not sector specific. Labour is mobile in most of these models, with workers receiving a common wage independent of the sector of employment. In this framework, wage gains should be experienced in all sectors as a result of free trade and the associated productivity gains. The empirical literature on the inter-industry wage structure provides evidence of persistent wage differentials across industries for workers that are otherwise ob-servationally equivalent.5 Explanations for these differentials include rent-sharing, unobserved skills, compensating differentials and efficiency wages. Including these kinds of considerations into the kind of trade model discussed previously allows for a more complex intra-industry wage structure. Furthermore, liberalization may also alter the interindustry wage as well as having effects on the average real wage. In what follows, I discuss the potential implications of liberalization when added complexity is introduced into the modelling of the labour market. 2.2.1 Rent-sharing In the presence of imperfect markets, it is possible for a union to extract rents from the firm through collective bargaining. Stylized models of union bargaining have been developed within the industrial relations literature and have been incor-porated in trade models such as those presented in Brander and Spencer (1988) and Mezzetti and Dinopoulos (1991). These models examine the implications of tariff policy when a unionized domestic duopolist competes with a foreign duopolist for the home market. When the tariff against imports rises, the home firm's market share and profitability improve. This in turn has implications for the wage and employment outcomes that the union can achieve through bargaining. In Brander and Spencer (1988), the interests of the home firm and the union coincide. In-5 For American evidence, see Katz and Summers (1989). For Canadian evidence, see Gera and Grenier (1994) 7 creasing the tariff raises both the profitability of the firm and the utility of the union. The union has preferences over both the wage rate and the level of employ-ment. As a result, the relationship between the tariff rate and the wage will depend largely on the willingness of the union to trade off employment for a higher wage. Mezzetti and Dinopoulos (1991) look at a similar model in which the home firm may relocate production to the foreign market and import back to the domestic market. In this case, the threat of relocation can be used by the firm to extract concessions from the union in the bargaining process. In this case, a tariff reduc-tion may actually be beneficial to the home firm, as it increases the profitability of relocation and forces the union to make larger concessions. The union is still hurt by tariff reductions. As with Brander and Spencer (1988), the exact relationship between the tariff and the wage rate will depend on the union's preferences over wage-employment combinations. When unions are present in an industry, they may affect the wages on workers with establishments that are not represented as well. Lewis (1963) argues that firms may pay workers a "premium" to deter unionization. In this case, policies which weaken the benefits of unionization will tend to reduce the likelihood of certification at non-union establishments. This in turn will reduce the need to pay higher wages to discourage unionization. Before continuing, it is worth summarizing the different effects that tariff re-ductions can have in the presence of unions. The first effect is on available rents; if tariff reductions increase import competition, then there will be fewer rents avail-able for the union to capture. Rents fall because demand for domestic production falls. This will result in decreased demand for labour as well. In this case, unless the union is willing and able to incur a large decrease in employment, the wage rate will fall. In the simple monopoly model of union behaviour, the union chooses a wage. The firm then hires according to its labour demand schedule. The union takes this into account when choosing the wage; as a result choosing the wage is equivalent to choosing employment and the wage simultaneously. Choosing to maintain the wage after a large decline in demand will result in large employment losses. The union could mitigate some of these losses by negotiating a lower wage and moving down the new labour demand curve.6 The second effect is on the abil-6Some models of collective bargaining allow bargaining over both the wage and employment. In these models, wage-employment combinations may lie above the labour demand curve and responses to product demand shocks can be complicated. See Gaston and Trefler (1995) for further details. 8 ity of the union to capture available rents. If tariff reductions make it feasible for firms to relocate to the trading partner and export back to the domestic market, this will affect the ability of the union to extract rents. Clearly the firm must be at least as profitable remaining in the domestic market as it would be by moving abroad, and all resultant bargaining outcomes will reflect this constraint. The third effect is on the non-union sector, through the deterrence effect. As long as unions offer workers substantial benefits, non-union firms may have to pay work-ers a premium to keep them from organizing. If trade liberalization reduces the benefits to unionization, then this premium will decline, which will reduce wages in the non-union sector. 2.2.2 Compensating differentials Another explanation for the wage differential is that there are differences in working conditions between industries. Industries with dangerous or arduous work-ing conditions would have to pay a higher wage to attract workers. Although a framework with workplace differences could generate a complex interindustry wage structure, it is not clear what the wage effect of trade liberalization would be. It is certainly not clear that there would be a clear relationship between the tariff of an industry and the wage rate in that industry. Suppose that good working conditions are a luxury good. In this case, if liberalization increases overall national income, the wage rate would tend to rise faster in industries with particularly bad condi-tions. However, there is no reason a priori to believe that there is a correlation between industries with the highest compensating differentials and industries with the highest tariffs. 2.2.3 Industry-specific human capital Workers may have industry specific human capital. Grossman (1983) presents a model in which capital is only partly mobile. In particular, capital becomes less productive when moved to a sector other than that for which it was originally intended. A case could be made that worker productivity behaves in a similar manner. Workers acquire skills on the job that make them more productive. This stock of skills, or human capital, may not be transferable to another industry. In this case, tariff policy may affect the relative return on human capital. However, workers with a large stock of such capital may still be better off working in the sector for which they have specific skills than working in a sector for which they 9 have no training. In this case, tariff changes can induce persistent wage effects. Even so, these effects should be diminished over time as young workers with little human capital move out of declining sectors and enter sectors with higher rates of returns. The exact relationship between the tariff structure and returns to skill may not be straightforward. The largest effects should be in those industries that contract as a result of trade liberalization. However, it may be the case that workers in different industries accumulate different amounts of industry specific capital. In this case, tariff reductions may well reduce the return to skills most in those sectors that faced the heaviest reductions. However, in sectors that were heavily impacted but where workers receive little training, one might expect little effect from tariffs on wages even if returns to human capital experience a large change. 2.2.4 Adjustment differentials Wage changes may also arise as a result of excess labour demand or supply. Workers may be attached to industries in the short-run. Workers may only change industries if conditions are unfavorable for a sustained period in time. In this case, changes in labour demand in an industry will not result in the migration of labour towards or away from that industry in the short-run. Instead, wages will tend to respond initially, as the labour supply is fairly inelastic. In this case, it may take some time before sustained wage differentials across industries result in the migration of labour between industries. 2.2.5 Summary of Predictions Based on the models outlined previously, it is unclear whether one would ex-pect to find a relationship between the tariffs applicable to an industry and the wages of workers in that industry. In a world with a mobile labour force, trade liberalization should have an economy wide effect on the wage. This effect may become quite complicated if considerations like compensating differentials are im-portant.7 Alternatively, if tariff reductions reduces the relative demand to sector specific human capital then such reductions may have a negative influence on the wages of workers in these sectors. A similar effect may be observed if freer trade makes it harder for workers to capture rents in the affected sector. 7It seems unlikely to me that the modest gains expected from C U S F T A would be sufficient to have much of an effect on compensating differentials. 10 2.3 Empirical Framework Following the existing literature on trade policy and wages, I assume that wage and employment outcomes result from either market clearing in a decentralized labour market, or from some form of collective bargaining. In either case, trade policy parameters are modelled as being product market determinants. In the case of decentralized markets, market clearing leads to reduced form equations of wage and employment levels for each industry i. Policy changes that increase product demand in the industry will tend to increase the level of employment, while the effect on wages will depend on the elasticity of labour supply. Freeman and Katz (1991) argue that in the presence of collective bargaining, wage and employment equations can be thought of as the outcome of a bargaining process. As noted in the previous section, the effect of policy parameters on the wage and employment outcomes will tend to be complicated, and will depend on the willingness of the union to trade employment for higher wages. The framework used here is similar to that used in Gaston and Trefler (1995). In particular, let the wages of individual i working in industry j at time t be given by ln(wijt) = Xitpt + DjtW*t + eijt (2.1) where Xjt is a vector of standard human-capital characteristics specific to individ-ual i, and Djt is a vector of indicator variables in which the j - th element corre-sponds to industry j. The first term reflects the extent to which an individual's wage can be explained by his or her own observable characteristics, while the sec-ond term is the portion of the individual's wage that can be accounted for by the industry of affiliation. Wages in an industry may change either because the returns to characteristics have changed on an economy wide-basis ( i.e. Bt ^ Pt+i) o r be-cause the industry specific component of the wage has changed (i.e. W*t f Wjt+i ). A more general specification might allow for the returns to some human-capital characteristics to depend on the industry of affiliation as well. For example, the wage equation would be written as ln(wijt) = Xit0t + D3tW;t(Xit) + (2.2) where Xit is a subset of human capital characteristics that have industry-specific value. Suppose that skills acquired on the job are firm or industry specific. In 11 this case, the return to these skills may be affected directly by changes in market conditions at the firm or industry level. To incorporate tariff policy into (2.1), I assume that the wage premia is deter-mined according to the process W*t = Yta + Zjt5 + Tjtl + Uj + ujt (2.3) where Yt is a set of controls for common cross-industry business cycle effects and other common observable shocks,8 Zjt is a set of controls for industry specific changes in labour demand and supply conditions, Tjt is a vector of trade policy parameters, p3 is an industry specific fixed effect, and Vjt is an industry specific shock. Industry characteristics potentially includes work conditions that result in compensating differentials and the nature of domestic competition, which affects the rents available for workers to share. The parameter of interest is the tariff effect, 7. In this specification, tariffs are constrained to have an instantaneous and long-lasting effect. Industry fixed effects are included to allow the industry pre-mia to depend on unobserved features of industries such as workplace conditions that are not directly influenced by trade policy and that have remained relatively constant over time. The common time effects control for business cycle effects. It also includes a component of wage growth that is common across all industries. Identification of the tariff effect is based on the extent to which tariffs and other industry-level variables can account for changes in the relative wages between in-dustries over time. In this specification, the wage rate in each industry is partly influenced by some common pressures over time. Additional variation in the wage rate between industries results from a combination of industry fixed effects, observ-able changes such as tariff reduction at the industry level, and unobserved shocks at the industry level. This specification is consistent with the kinds of models that were laid out in the previous chapter.9 In this paper, the tariff policy vector will consist of the own industry American and Canadian tariffs. In principle, tariff changes in other industries could result in substitution effects. Following the literature I ignore such substitution effects. This is roughly equivalent to assuming that cross elasticities between American and Canadian goods in the same industry are much larger than those elasticities 8Common time effects are included as part of the industry premium for expositional ease. 9In particular, it is consistent with a general equilibrium model where worker migration keeps the wage rates from "drifting" apart too much. 12 between goods in different industries. In the previous section, it was noted that trade liberalization also had the potential to have a common effect on wages across industries. This would be equivalent to making the returns to characteristics de-pend on some summary measure of the agreement. This issue is not addressed in this paper.10 Estimates of the tariff effect are established in a two step process. In the first step, equation (2.1) is estimated using the data from the labour surveys. The estimates of the industry-year effects are then carried over to the second stage, where they are used as the dependent variable in estimating equation (2.3).11 The population weights included with the labour survey are utilized in the first stage regression. In the second stage, I experiment with several weighting schemes. 2.3.1 Existing Empirical Studies There are three published papers that have attempted to assess the impact of CUSFTA on wage and employment outcomes in Canada. The first of these studies, Gaston and Trefler (1997), uses annual industry averages of employment and weekly earnings. This data is based on Statistics Canada surveys at the establishment level and spans the period 1980-93. As this data is not at the level of individual workers, the study involves estimating first-differenced versions of equation (2.3), where the dependent variables are industry level employment and earnings. Gaston and Trefler find little evidence that earnings growth is adversely impacted by tariff reductions. The is some evidence that employment contracted more in those industries that faced the largest tariff reductions. However, according to Gaston and Trefler, most of the job losses occurring in the post-CUSFTA period were the result of the recession that began in 1990; in their framework, only 9-14% of job losses that occurred after 1988 can be accounted for by the tariff reduction. The second paper, by Beaulieu (2000), uses more disaggregate data to repeat the analysis performed in Gaston and Trefler (1997). Bealieu uses a data set which spans the period 1983-1996 and includes separate series for production and non-production workers. These groups of workers are used as proxies for unskilled and 10It is not clear what the appropriate statistic to use would be to capture the general equilibrium effect of CUSFTA on the wage wage rate. Beaulieu (2000) considers the average tariff rate. He finds no evidence that average wage movements are related to this variable. 1 1 To be more precise, wages are predicted for a worker type. Since predictions are based on a common set of worker characteristics, all differences in wages must be attributable to industry and year effects. 13 skilled workers, respectively. Beaulieu finds little evidence that the tariff reductions had an adverse effect on the earnings of either type of worker. Employment losses are found to occur mostly among production workers; as a result of these losses the skill content in impacted industries increased. In the third paper, Trefler (2001) re-examines the wage-employment issue as part of a comprehensive study of CUSFTA using data from the Annual Survey of Manufactures (ASM). This data allows Trefler to use more disaggregate industry categories than were available in earlier studies while working with worker cat-egories similar to those used in Beaulieu (2000). He finds evidence that labour productivity and the relative wages of production workers increased in the sectors most impacted by tariff reductions. Tariff reductions resulted in a modest 5% wage increase in the most impacted industries. Gaston and Trefler (1995) use cross-sectional micro data to estimate a similar model, using the two-step procedure outlined in the previous section . In their paper, they use one year of the Current Population Survey (CPS) data to inves-tigate the relationship between union wages and tariff and non-tariff measures of import protection. The CPS surveys individual workers, and includes information on the demographic characteristics of individual workers, along with information on the wage, union status and industry of affiliation of the "main" job held by each worker. This data is combined with detailed information on tariff and non-tariff barriers. Gaston and Trefler find that imports and tariff protection are important determinants of the union wage premium. Lower tariffs and lower imports are both associated with a higher wage. Although the tariff result is counter-intuitive, it is robust to an instrumental variables specification. The authors conclude that the cross-sectional nature of their data may be insufficient to fully disentangle the tariff effect on wages. 2.4 Data My labour data is drawn from a collection of cross-sectional and longitudinal data sets that span the period 1981-98. In particular, I use the 1981 Survey of Work History (SWH), the 1984 Survey of Union Membership (SUM), the 1986-87 Labour Market Activity Survey (LMAS), the 1988-90 LMAS, the 1995 Survey of Work Arrangements (SWA), and the 1997 and 1998 Labour Force Survey (LFS). The year of each survey corresponds to the year for which the questions actually 14 apply. A l l surveys are either from the LFS or supplements to the LFS and are comparable in design and construction.12 Each survey is meant to be roughly representative of the Canadian population. In all of the analysis that follows, I use the sample weights included to reflect non-random sampling used in the construction of the surveys. Although the LMAS follows individuals for several years, I treat these data sets as cross-sections. For each survey year, I retain those individuals aged 16-6913 that are paid workers and were members of the labour force during November of the year of the survey. I examine the usual hourly wage from the main job held in that month, with the main job defined as the one with the most hours per week. I do not include overtime wages in my analysis. A l l wages are converted to 1998 dollars using the CPI. In January of 1990, the LFS changed the way in which educational information was collected. Prior to 1990, individuals that did not finish high school but completed a post-secondary diploma were grouped with those individuals with some high school education. After 1990 these individuals are included with those individuals that completed both high school and a post-secondary diploma or certificate. As a result of these changes, my educational groups are not perfectly comparable before and after 1990. An analysis of the educational attainment of the synthetic cohort of men that were aged 25-34 in 1981 suggests that the educational attainment within this group has remained constant across the classification schemes.14 However, the fact of this seam in the data should be kept in mind. In each survey year, for each job held, the individual is asked to identify the industry in which he is employed. On the public use tapes, industry information is 1 2One exception to this is that the SWH records whether individuals were union members but not whether they were covered by a union contract if they were not members. All the subsequent surveys recorded union coverage and membership. I believe that the correct manner in which to capture union effects is to measure their impacts on all workers covered by their contracts, whether they or not they are union members. To maintain consistency, I considered dropping the SWH from my analysis, but 1981 is a crucial year in that it is cyclically similar to 1988 and 1998, a feature that will be exploited later in the analysis. Instead, I use 1984 data to estimate a probit on non-union members who are covered by union contracts. I use the estimated coefficients from this exercise, along with individual covariate values and draws from a standard normal distribution to form fitted values of an underlying index function corresponding to union coverage. Those with index values greater than 0 are assigned as covered by a union contract. This selects particular individuals while preserving predicted probabilities of coverage for sub-groups defined by common covariate values. 1 3I discard those individuals aged 16-19 claiming to have at least some post-secondary education and those individuals aged 16-24 claiming to have a university degree. 14These results are available upon request. 15 reported at the 2 digit SIC level. 1 5 For most years, industry categories in mining and manufacturing match up with the 22 industries used in Gaston and Trefler (1997) and the 19 manufacturing firms used in Beaulieu (2000). The last two years of data are an exception, as the miscellaneous category is expanded to include tobacco products, leather and coal and petroleum products. For the purpose of compatibility across years, these industries are merged in previous years. Matching Canadian and American tariff series are created by forming annual averages over the appropriate tariffs in which the 1987 import and export volumes are used as weights.16 The tariff and American employment series are from the same sources as those used in Gaston and Trefler (1997) and Beaulieu(2000).17 In particular, post-1993 figures for American employment are from the Earnings and Employment series available from the Bureau of Labour Statistics (BLS) and are reported for all workers. Tariffs for the years 1994-98 are from Magun et al. (1988). Where appro-priate, the weighting scheme described above is used to form the new miscellaneous category. 2.5 Empir ical results Recall that previous studies using aggregate data found that the CUSFTA tariff reductions had either no effect or a small positive effect on the wages of workers in impacted industries. The use of aggregate data may have hidden a tariff effect on wages, if either (i) the distribution of worker characteristics differed substantially between high and low tariff industries and tariff reductions has differing effects on the returns to different characteristics or, (ii) if the composition of characteristics changed in an offsetting manner in the affected industries. To illustrate the first case, suppose that high tariff industries employed a greater proportion of female workers (as I will show to be the case). A number of studies suggest the the male/female wage differential closed marginally over the period of study. In this case, in the absence of tariff effects, the mean wage in high tariff industries would have tended to improve relative to the mean wage in low tariff industries. If tariff 1 5In 1981, industry classifications were according to the 1970 SIC. This was changed to the 1980 SIC for the remaining years in the data used here. The 1980 revision was reasonably minor, suggesting that the industry groups for 1981 and the remaining years are comparable. 1 6This weighting scheme is consistent with the scheme by which the post-1987 tariff series were created. See Magun et. al. (1987). 17Gaston and Trefler have graciously made this data available at the NBER web site 16 reductions worked to reduce the wage, then persistent compositional differences in the workforce, combined with different secular wage trends between different types of workers would work in opposite directions on the mean wage. The second case is more subtle, and may include trade induced changes in worker composition within industries.18 Suppose for example that the tariff reductions reduced the relative demand for less skilled workers in the impacted industries. In this case, trade liberalization would result in an employment shift towards more skilled workers. Even if the wages of both skilled and unskilled workers were negatively impacted, the ensuing compositional change might be sufficient to result in a higher mean wage in the industry. To the extent that production and non-production worker categories proxy skill categories, both Trefler (2001) and Beaulieu (2000) controlled for possible compo-sitional bias relating to shifts in the returns to different skill groups and the returns to these groups. However, it may be the case the other kinds of compositional ef-fects related to gender, unionization, age and tenure distributions may have been important contributors to wage developments before and after CUSFTA. Further-more, production and non-production workers may not be good proxies for skill if the complexity of the production process varies across industries. Some industries may require highly skilled production workers. By having a measure of the educa-tion attainment of individual workers, I overcome this problem. In what follows, I first assess the extent to which workforce composition varied across industries. No attempt is made to attribute any shifts to CUSFTA. I then proceed to use the regression framework outlined in section 2.3 to assess the extent to which the CUSFTA tariff reductions impacted the wages of different worker types. 2.5.1 Wage and tariffs over the period 1981-1998 I begin with a brief description of the broad wage and tariff developments during the study period. Prior to CUSFTA, Canada and the U.S. tended to provide tariff protection to the same set of industries. Furthermore, over the pre-CUSFTA period of 1980-88, tariffs were being reduced in a number of industries in accordance with agreements reached during the Tokyo Round of GATT. Table 2.1 presents the Canadian tariff against American imports and the log mean wage for each industry 1 8The issue of whether the tariff reductions affected the composition of workers within indus-tries is deferred until the third chapter of this thesis. For now I am interested only in the effect of tariff reductions on the wage of workers with particular characteristics 17 in my data set in 1981, along with the changes in each of these series in the pre-and post-CUSFTA periods. The series are sorted in ascending order based on the mean wage in 1981. The industries that received the greatest protection tended to be low wage industries. The Tokyo Round reductions tended to be largest for industries with moderate rather than high level of protection. In contrast, CUSFTA involved eliminating tariffs across all industries; as a result, the largest cuts were experienced by high tariff industries. To further illustrate the nature of the tariff cuts, the annual average tariff rates are plotted in figure 2.1. As can be seen, tariff rates were declining over the period 1981-88. Further cuts occur at an accelerated pace over the period 1988-93, with further cuts occurring over at a slower rate over the period 1993-98. The kinked nature of the cuts over the period 1988-98 is a feature of the CUSFTA agreement; depending on the commodity, tariffs were phased out over a one, five, or ten year period beginning in 1989. By 1993, tariffs had been reduced to zero for a number of commodities. Figure 2.2 shows a plot of the American and Canadian tariff for each industry in 1988, along with a line obtained by a simple regression of the American rate on the Canadian tariff.19 There is a strong relationship between the two series. Furthermore, the tariff rates tended to be reduced simultaneous. As a result, this relationship tends to be preserved over the entire study period. 2 0 This multi-collinearity suggests that may not be possible to identify separately the effect on the Canadian and the American tariff. An examination of the wage series suggests that wages grew modestly over the pre-CUSFTA period; there is little evidence that the pattern differed greatly based on either the 1981 level of the tariff or the magnitude of the Tokyo round cut. In contrast, wages over the post-CUSFTA period tended to decline. Once more, there is no obvious relationship between the magnitude of the change in the wage and the change in the tariff across industries. 2.5.2 Workforce Characteristics by Industry I address the issue of compositional change by examining the distribution of worker characteristics in low, medium and high tariff industries. In Table 2.2, the mean characteristics of workers in my panel of "tradeable" industries are presented for the years 1981, 1988, and 1998. These years were chosen because they represent 1 9The estimated slope coefficient is .75. The R2 is .80. 2 0The correlation coefficient between the two series is .87. 18 similar points on the business cycle. Furthermore, 1988 is the year immediately prior to the implementation of CUSFTA, while 1998 represents the first year in which the tariff reductions were complete. The breakdown includes the overall gen-der ratio and union coverage rates, along with breakdowns within gender groupings by age and level of education. Taken as a whole, the "tradables" sector is highly unionized and employs mostly men. The gender ratios have been remarkably con-stant over the study period. Union coverage has slipped somewhat, particularly between 1988 and 1998. The skills content of the workforce, as measured by ed-ucational attainment, has been gradually increasing. Furthermore, the workforce has been gradually aging; this pattern likely reflects demographic patterns within the general population of Canada. At a more disaggregate level, industry distributions of worker characteristics are presented in Tables 2.3 and 2.4. In each table, industries are sorted in descending order according to the 1988 tariff rate against American imports. Table 2.3 shows the proportion of workers covered by a union contract and the proportion of workers that were male in each industry for the years 1981, 1988, and 1998. The last three rows report the same statistics for the unweighed mean of the six industries with the highest and lowest tariffs, along with the mean for the seven industries with medium tariffs in 1988. Two trends are suggested by the last three rows. First, the most protected industries historically have lower coverage rates on average than the least protected industries. Furthermore, over the study period, union coverage has declined across all industries, with most of the decline occurring in the 1990s. The decline seems to be somewhat more pronounced in the high tariff sector 2 1 . Second, industries that have traditionally received protection also employ a higher proportion of women. This result is partly skewed by the inclusion of the clothing and textile industries, which predominantly employ women. Even in narrow industry categories the gender breakdown has remained largely unchanged throughout the study period. Patterns of employment based on age and skill level are examined in Table 2.4. Unskilled workers are defined as those that have completed no educational certification beyond a high school diploma. Young workers are between 16-34 years of age. The last three rows indicate that whereas skill-upgrading has been occurring in all industries, it is those industries that were the most protected prior to CUSFTA that have and continue to make the heaviest use of unskilled 2 1The issue of whether or not unionization was affected by CUSFTA is deferred to Chapter 3. 19 labourers. Skill-upgrading appears to be common to all industries. There is no obvious relationship between protection and the age of the workforce. The average age or the workforce has been increasing in all industries. The employment patterns described above have conflicting implications for wage movements within and across manufacturing industries. In particular, the combination of skill-upgrading and an ageing, more experienced workforce should result in higher wages in all industries. However, as these two trends have been oc-curring in all industries, there is no reason to expect that they should have resulted in diverging wage paths across high and low tariff industries. In contrast, declining coverage rates across industries should result in lower mean wages. Econometric studies of union wage effects suggest that union status results in a 15 % wage pre-mium after controlling for other characteristics.22 Given that union rates dropped off somewhat more rapidly in the high tariff sector over the period 1988-98, the wages in these sector should have fallen relative to those in the low tariff sector. 2.5.3 Regression results To estimate the equations outlined above, I split the data set into well-defined subgroups based on union status and education. As a result, the tariff effect is allowed to vary across these worker groupings. The relative return to a number of characteristics such as gender and tenure are allowed to vary from year to year. In principle, changes to these returns could be caused by trade liberalization as well. However, I assume that all workers within broad union and skill categories face the same effect. This effect is limited to the impacted industry. Further breakdowns in the groupings based on alternative measures of experience are explored later to address whether or not liberalization affected the returns to human capital. No attempt is made to link CUSFTA to economy wide changes in the Canadian wage structure 2 3 . The discussion in section 2.2 suggests that tariff reductions may have impacted 2 2For a recent survey on union effects, see Kuhn (1998). 2 3In the U.S.A., a number of researchers have examined the extent to which increased trade lib-eralization can account for the increasing wage differential between more and less skilled workers. Beaudry and Green (2000) argue that in Canada, the most striking growth in wage inequality has occurred across cohorts. More recent job market entrants face a depressed lifetime wage profile relative to earlier entrant cohorts. Research by Morissette (1997a) indicates that changes in sectoral employment across cohorts cannot account for this decline. More recent research by Green and Townsend (2002), presented as Chapter 4 of this thesis, indicates that some of the cohort inequality can be accounted for by declining union representation among more recent cohorts of job starters. 20 worker wages by either changing the rents available through collective bargaining or by reducing the returns to industry specific human capital. As a starting point, I estimate separate wage equations for four groups of workers based on union status and skill. As defined previously, less skilled workers are those workers that have obtained no more than a high school diploma. More skilled workers include those workers with some post secondary training and those with a university degree. The union sector includes all workers that are covered by a collective bargaining agreement. Table 2.5 presents the second stage estimates for the less skilled workers in the non-union sector.24 The results in column (1) are based on a specification in which the industry premia is a function of the applicable American and Canadian tariffs, the log of American employment in the corresponding industry, and a set of year dummies and industry dummies. Following Gaston and Trefler (1997), I include the American employment series to control from broad trends in employment relating to de-industrialization, technology and trade related effects unrelated to CUSFTA (i.e. shifting of production of low end manufactures to developing economies). Given the relative size to the two economies, Gaston and Trefler (1997) argue that American labour market outcomes were not affected by CUSFTA. They formally test the employment series for exogeneity and.are unable to reject the null hy-pothesis. Time dummies are used to control for business cycle effects and broad cross-industry movements in the returns to skills. This specification is similar to the specifications used in Gaston and Trefler (1997) and Beaulieu (2000). How-ever, instead of using macro-variables and a time trend to control for business cycle effects and secular wage trends, I use year fixed effects. This allows for more flex-ibility at the cost of estimating additional variables. An alternative specification is considered below in which the year effects are replaced with a parametric spec-ification depending on macro-variables similar to those used in previous studies. The Canadian tariff coefficient takes on a value of .007, which, given the log form of the wage, indicates that a one percent reduction in the statutory tariff rate reduced the hourly wage of a worker in the impacted sector by approximately 0.7%. This is a large direct effect. In the clothing industry, for example, the Canadian tariff against American imports in 1988 was 17.2%. The point estimate indicates that for less-skilled, non-union workers in this industry, tariff reductions resulted 2 4Given the overwhelming number of coefficients estimated, results are presented in an ap-pendix to my dissertation and are available upon request. 21 in a 12.4% wage reduction relative to workers in industries with no tariff decrease. Although this effect seems large, the tariff rates for most industries were small. The average tariff rate prior to CUSFTA was 4.5%; the impact of CUSFTA on the "average" worker would thus be a wage reduction of around 3%. To check the robustness of this result, I consider several alternative specifica-tions. Column (2) addresses the issue of multi-collinearity between the two tariffs by dropping the American tariff. As would be expected in the presence of strong correlation, dropping the American tariff results in a lower standard error of the estimate. The resulting estimate increases slightly. In column (3), American em-ployment is dropped as well; there is no noticeable effect on the tariff estimate. In column (4), the results of specification (2) are reported with standard errors that have been corrected for potential heteroskedasticity using White's method. The industry-year effects from the first stage regression are estimated with varying de-grees of precision, which suggests the results may be driven by one or more "odd" values. The specification in column (5) is identical to that of (2), but each obser-vation is weighted by the inverse of the standard error of the first stage estimate; this places more weight on those observations that are more precise. Finally, in column (6), the year effects are replaced with a series of covariates to control for business cycle effects and cross-industry wage trends.25 This specification is sim-ilar to those used in previous studies. The advantage of the resulting regression equation is that fewer parameters need to be estimated. The disadvantage is that the common time component of the wage is much less flexible. These alternative specifications have little impact on the tariff coefficient. In specifications in which the American tariff is dropped, I interpret the Canadian tariff coefficient as being the "bundled" effect of the Canadian tariff reduction and the reciprocal cut to the equivalent American tariff. The results presented here suggest that for impacted workers in the low skill, non-union grouping, a one point decrease in the statutory Canadian tariff, along with the corresponding reduction to the American tariff, resulted in a 0.7% reduction in the wage rate. Similar regressions were estimated for each of the remaining groups. A sum-mary of the results are presented in Table 2.6. The results are based on a model in which the second stage regressors are limited to the Canadian tariff and year and industry dummies (column (3) in Table 2.5). These results suggest that there 2 5 I n particular, I include measures of the annual unemployment rate, the exchange rate, and a quadratic in time. 22 was a negative wage effect resulting from CUSFTA for workers not covered by a collective bargaining agreement, regardless of skill level. In the union sector, a smaller negative effect was experienced for less skilled men, while more skilled men were unaffected. Results were generally robust to the specifications outlined in the previous paragraph. The largest change results from including the American tariff in the specification for the group of less-skilled, unionized workers. In this case, the Canadian tariff coefficient becomes insignificant, while the American coefficient is .010.26 As Figures 2.1 and 2.2 show, the Canadian tariffs were generally higher than the American tariffs. Given the collinearity between the two tariff series and the relative magnitudes of the statutory rates, these results should not be taken as being at odds with the results of the preferred specification. The results indicate that the wage in the union sector was less responsive to the tariff than the wage in the non-union sector. This seems at odds with a model in which the industry premia reflect rent-sharing; CUSFTA was expected to increase competition, which in turn should have reduced available rents. Collective bargaining is the most obvious way for workers to obtain a share of these rents. In this case, union workers should have been more affected by liberalization. However, it may be the case that unions opted to trade off employment to maintain wages, in which case one might not expect to observe a wage effect. By the same token, the wages of non-union workers may contain a threat premium that firms pay to deter unionization. If liberalization reduced the perceived benefits of unionization, then the threat premium would fall in a manner consistent with the results presented in Table 2.6. A more thorough investigation of these possibilities would require an examination of employment adjustments in the two sectors. An alternative explanation for the patterns is that the demand for industry-specific human capital were negatively affected by liberalization, but than union bargaining mitigated this effect. The existing literature on unions suggests that wage structure in the union sector is more compressed than that found in the non-union sector.27 This compression occurs across dimensions commonly associated with human capital accumulation, such as age or experience. As a result, union wages may not be strongly correlated with an individual's own skills and hence pressures that alter the demands for certain characteristics may not have a similar effect in the union sector as it has in the non-union sector. 2 6 The standard error of the estimate was .003. 2 7See Kuhn (1998) for a summary of findings. 23 To address the extent to which the tariff reductions reduced the returns to human capital, I follow Gaston and Trefler (1995) and split non-union workers up further into groups with less and more human capital. I then re-estimate the tariff effect for each group to determine whether the tariff effect differed according to human capital accumulation. Two alternative measures are used to proxy for job specific human capital. The first is based on job tenure; I divide workers into those with less than and more than five years of tenure in the current job. The second is based on age; I estimate tariff effects separately for those aged 16-34 and those aged 35-64. The estimated tariff impacts for inexperienced and experienced workers are presented in Table 2.7. The results for less educated workers are comparable to those presented previously. It should be noted that the tariff effect for more experienced workers tends to be larger than that for less experienced workers, though these differences are not statistically significant. The results for more educated workers are mixed. Using job tenure to proxy for experience results in a significant and positive tariff coefficient for less experienced workers and no effect for more experienced workers. In contrast, using age as a proxy for experience results in no effect on less experienced workers and a positive and significant tariff coefficient for more experienced workers. A potential problem with this kind of analysis is that the sample size for each industry year effect becomes quite small. For example, for the more educated group of workers in the analysis of Table 2.7 , a number of industry-year effects in the first stage regression are based on a single observation once the data is split up according to experience. This suggests that the data set may not be sufficiently large to properly untangle a returns to human capital effect of tariffs for a number of sub-groupings. The results for less-skilled, non-union workers, where the first stage sample sizes are larger, suggest that returns to human capital may be accounting for the tariff effects observed in Table 2.6. 2.5.4 Past and Present Results: A Reconciliation Unlike previous studies, I find evidence that the CUSFTA reductions had a negative impact on the wages of impacted workers. Gaston and Trefler (1997) and Beaulieu (2000) both found that the tariff reductions had both an economically and statistically insignificant impact on the wage. Trefler (2000) found that the 24 CUSFTA cuts raised wages by 5% in the most impacted sectors.28 As I argued earlier, micro data allows the researcher to control from a number of compositional issues that may potential obscure detected a wage impact from the CUSFTA liber-alization. Unlike previous researchers, I find that reduced protected impacted the wages of workers in the affected sector. One possibility is that the kinds of compo-sitional effects described above were indeed important. Another, more troubling possibility is that the difference in results is a result of different earnings concepts or other methodological differences between the two surveys. In this section, I address each of these issues. I begin by examining the comparability of the two earnings series. The earning measure from the payroll data used in previous studies is a measure of weekly average earnings. This measure is not a pure measure of the price of labour, as it contains a quantity measure (weekly hours) The measure from the labour data used here is the hourly wage rate, and is, I argue, a truer measure of the price of labour inputs. Conceptual issues aside, though, the easiest way to determine the importance of controlling for worker characteristics in uncovering a tariff effect on wages is to proceed to estimate a wage equation without first exploiting the micro dimensions of the data set. To do this, I construct average hourly earnings in each industry-year cell using the sample weights from the surveys. I then estimate a simple model of the wage on the Canadian tariff, along with industry and year effects, using these averages. The Canadian tariff co-efficient in this specification is .002, with a standard error of .002. The conclusion from this specification would be similar to that found in previous studies: tariff reductions had no effect on earnings in the affected sectors. To further examine the results, I re-estimate the results for non-union, less-educated workers (Table 2.5). I begin with a specification in which I construct raw means (using the weights) for each industry-year cell and proceed directly to the second stage regression. Results are presented in column (1) of Table 2.8. I then estimate a version of equation 2.1 in which human capital characteristics are restricted to having the same effect in each year (i.e. Bt = [3). Finally, column (3) presents the result established previously. The results are striking. By not further exploiting the micro-dimensions of the data, I find a negligible tariff effect (column 2 8Trefler's data set permits him to look at more disaggregate industry categories than are available here. The most impacted industries consists of the 71 industries that underwent the largest cuts. Cuts ranged from between 5% and 33% for these industries. At the industry level wage gains range from 2.5% to 16.5%. 25 (1)). Inclusion of time-invariant returns to human capital yields a weakly signifi-cant and positive wage effect of protection (column(2)). The preferred specification (column 3), which includes time-variant returns to human capital characteristics, suggests an even larger effect of tariff reductions. As the previous paragraph shows, I produce results that are very similar to two of the previous studies when I exploit the same information that was used in these papers to estimate a similar relationship. Exploiting only the industry variation leads me to the same conclusions reached in Gaston and Trefler (1997) and Beaulieu (2000); tariff reductions had no appreciable effect on the wage structure. This similar result is generated despite the use of different earnings measures. Including worker characteristics but fixing the returns to characteristics generates a small negative impact of tariff reduction. This specification amounts to controlling for compositional changes.29 Controls for composition are sufficient to generate results similar to those of the prefered results. Allowing prices of characteristics to vary produces an even larger tariff effect, and suggests that these kinds of wage dynamics were having a differential impact on industries in a way that varied systematically across industries. These results suggest that both compositional effects and changes to the returns to characteristics may have had wage implications that at the industry level largely offset negative wage impacts arising from CUSFTA. To look at this issue more fully, I examine how compositional and price changes varied across groupings of industries based on the 1988 tariff. In particular, based on equation (2.1), the mean log wage Wjt, of workers employed in industry j at time t is given by K wjt = £ XktBtwk + W*t (2.4) fe=i where K is the set of workers in the sample in industry j at time t and Wk is the (normalized) sample weight of worker k. The difference in the mean wage of industry j between period t and i is thus a combination of changes in the samples (i.e. worker characteristics) between periods, changes in the returns to those characteristcs and changes in the industry effect. Suppose that Xkt can be partitioned into two sets of characteristics, Ykt and Zkt. Then the difference in the 2 9If prices to characteristics were changing over time, then the industry-year effects estimated in the first stage would no longer be pure industry effects, but would also contain these wage movements in a manner that is weighted according to the composition of the workforce in each industry. 26 mean industry wage between periods t and t is given by K k K k Wjt-Wjt = (J2 Yktatwk - Y YktaiWk) + (Y zktltwk - Y zktllwk) + (w*t - WT{) k=l fc=l k=l k=l (2.5) where the first term in brackets is the difference between periods attributable to changes in the distribution and return to characteristic y, the second term is the difference between periods attributable to changes in the distribution and return to characteristic z and the third term is the difference attributable to a pure industry effect. Table 2.9 presents results based on a decomposition of this type, for the sample underlying the results of table 2.5. Five sets of characteristics are used, based on the categorical data on gender, education, age, job tenure, and province of residence used in the first stage regression. Differences between 1988-1981 and 1998-1988 are broken down for each set of characteristics, along with the residual that is attributed to an industry effect. Total changes are also presented. As before, results at the industry level are averaged into three categories based on the 1988 Canadian tariff. During the period 1981-88, wages move largely in tandem regardless of the 1988 Canadian tariff. This can be seen from both the "Industry" and "Total" Rows. The relative wage gaps closes between low and high tariff industries. This appears to be largely an industry effect, though it is notable that the gender effect actually causes the average relative wage in high tariff industries to increase relative to the low tariff category. Figure 2.3 plots the "pure" effect of being male over the period. From 1981 to 1988, this premium increased, and should have had a larger effect in the male dominated low tariff industries. Yet this does not appear to be the case, suggesting instead that gender-ratios must have changed in an offsetting manner. Table 2.3 provides some evidence for this, as the percentage of the workforce in the high tariff sector that was male increased by five percentage points, while the same figure fell by 2 percentage points in the other two groups. During the period 1988-98, the wage structure starts to spread, as the "In-dustry" and "Total" rows indicate. The overall spread is much smaller than the "Industry" spread, suggesting that human capital characteristics are partly off-setting whatever is changing the "pure" inter-industry wage structure.30 Gender 3 0 Clearly the hypothesis here is that CUSFTA is playing a large role in changing the wage 27 differentials play a large role in the offsetting effect. As the composition was pretty static from 1988 to 1998 (see Table 2.3), this difference is likely a result of the combination of the gender gaps closing, as depicted in figure 2.3, and the greater proportion of women employed in the high tariff industries. Furthermore, provincial and age effects play equally important roles. The role of provincial ef-fects likely stems from the fact that different industries have different geographical locations. As a result, wages will be subject to economic pressures at the provin-cial level. The finding that the age distribution results in differences in groups is less clear. The tabulations in Table 2.2 suggest that the three groups had similar age structures. A complete understanding of the role of the age structure as an offsetting force would require a further decomposition into price and quantity ef-fects, using a greater level of disaggregation than my simple distinction between "young" and "old" workers. 2.6 Interpreting the results In section 2.2, several theories were presented that indicated that CUSFTA might have an impact on the inter-industry wage structure. In this sub-section, I examine the extent to which my findings are consistent with these theories. 2.6.1 Collective bargaining and rent-sharing The union wage premium is thought to partly reflect rent-sharing between firms and workers. Changes in the environment that affect the profitability of the firm will also impact wages and employment in an unionized workplace. CUSFTA, which involved both the wide-scale dismantling of protection to domestic produc-ers and increased access to the American market, certainly had the potential to alter the profitability of domestic production units. In addition, even in the ab-sence of large impacts on profitability, CUSFTA opened up the potential to shift production to the U.S. while maintaining access to the Canadian market. The threat of relocation could then potentially be used as leverage by management in establishing the terms of new collective bargains. Despite a body of theoretical work that suggested that changes in tariffs could alter the wages of unionized workers, I find little evidence that CUSFTA changed the wage structure in the union sector. Instead, most of the impact appears to be structure. 28 in the non-union sector. Furthermore, the stability of the union wage structure, combined with the relative decline in wages for non-union workers in impacted industries, indicates the the intra-industry union premium increased in those in-dustries most impacted by CUSFTA. The impact of tariff reductions on union wages through changes in product market conditions is ambiguous, depending on the union's willingness to trade employment for high wages. In principle, decreased demand could lead to lower employment by higher wages. The threat of outsourc-ing, on the other hand, was unambiguously expected to drive down union wages. This suggests that while CUSFTA may have reduced available rents, it did not affect the ability of unions to capture a share of available rents through collective bargaining. The wages of non-union workers could also have been impacted if CUSFTA reduced the benefit to unionization. I argue above that there is little evidence that CUSFTA prevented unions from capturing rents. However, available rents may have decreased in impacted sectors. Declining rents would be reflected in the non-union wage if part of the wage reflects a premium paid to deter unionization. The observed wage patterns are consistent with a world in which trade liberal-ization reduced available rents. Unionized work places responded by accepting cuts in employment to maintain wages, while non-union workers received lower wages because lower benefits to unionization reduced the premium needed to deter unionization. A decline in rents would almost certainly be related to a decline in product demand, which would also lead to decreased labour demand. As I outline in the next section, decreased labour demand may be sufficient to account for the observed decrease in wages in the non-union sector. 2.6.2 Specific human capital Gaston and Trefler (1997) and Beaulieu (2000) both found that employment decreased disproportionately in the 1990's in those sectors most heavily impacted by tariff reductions. This suggests that labour demand was adversely affected in these sectors. With competitive labour markets, one would not expect this shift to result in a sustained change in the intra-industry wage structure; workers should simply migrate away from the low demand sectors until no arbitrage potential ex-ists. Re-allocation may have a lasting impact on the relative wages of different worker types, based on skills or some other dimension other than industry of af-filiation. As noted in section 2.2, this situation changes if workers have human 29 capital specific to an industry. In this case, the rate of return on the capital may decrease due to a tariff reductions. However, as the capital receives no return in other industries, the worker may choose to remain in a diminished industry. Wage differentials will gradually dissipate as older workers with substantial stocks of human capital retire and newer workers opt to seek employment in alternative sectors. I find some support that wage losses experienced in the non-union sector are consistent with a specific human capital model. The tariff impact is higher for less-educated workers with more experience. This holds whether age or job tenure are used to proxy for experience. This is consistent with the notion that those workers with a larger accumulation of skills experienced a more adverse impact from tariff reduction.31 Results were less clear for more-educated non-union workers, where the results depended on the proxy for human capital that was used. However, there were doubts about the reliability of these results, given the small sample sizes involved by splitting the data up into smaller groups. If the wage losses in the non-union sector were indeed driven by decreased demand for certain types of human capital, how were union workers able to avoid a wage loss? A decrease in labour demand, presumably driven by increased product market competition, should have impacted the collective bargaining outcome as well. One possibility is that it did; unions responded differently by maintaining the wage and allowing job losses among workers without seniority. In principle, this hypothesis could be tested by examining employment and job tenure in the union and non-union sectors before and after CUSFTA. Given the small sample size of the data used here, such an exercise is not feasible.32 2.7 Summary and Conclusions The estimates that I have obtained using micro data indicate that the tariff structure has a substantial impact on the relative wage structure of manufacturing industries. In particular, for workers in impacted industries that are not covered 3 1This is actually true of the basic log linear specification as well. A component of a worker's wage results from his human capital. A one percent decrease in earnings brought about by tariff reductions will involve a one percent decrease on each unit of human capital that a worker holds. As a result, more experienced workers stand to lose more from tariff reduction. 3 2 Although the included weights should sum to the total labour force, counts are not always comparable across years. This occurs because Statistics Canada periodically re-bases the weights using Census results. 30 by a union contract, a 1% reduction in the statutory tariff rate results in a .7% reduction in the relative wage. This effect is independent of the skill-level of workers. Prior to CUSFTA, tariffs on the order of 10% were not uncommon; my estimates suggest that eliminating tariffs in these industries resulted in a 7% wage loss relative to similar workers in industries that had low tariffs prior to 1988. Results for the union sector were less severe, with less-skilled workers facing a .4% reduction in wages per percentage point reduction in the relevant tariff. The wages of skilled workers in the union sector were unaffected. It seems likely that labour demand was also impacted in the union sector, but that contract bargaining allowed unions to trade off employment for wage maintenance. Tariff policy could have a persistent impact on the inter-industry wage structure for several reasons. Two possibilities considered here are rent-sharing and returns to sector-specific human capital. The stronger impact in the non-union sector seems at odds with the notion that CUSFTA created outsourcing possibilities which eroded the ability of unions to capture rents. It may however have reduced the amount of rents available to be captured. As noted above, unions may have been able to avoid wage losses by agreeing to lower levels of employment instead. I attempted to address the extent to which the tariff impacts reflect returns to human capital. Unfortunately sample sizes for particular industries in particular years are often very small when the data set is divided up into groups of more and less experienced workers. Despite these limitations, results for non-union, less-skilled workers provide weak evidence that the impact of tariff reduction was stronger for more experienced workers in this group. The results for more skilled workers in the non-union sector are mixed and are highly sensitive to the variable used to proxy for experience. The results presented here suggest that tariff reductions can have large effects on the wages of those directly impacted. In the clothing and furniture industries, tariffs were 17.2% and 12.6%. These numbers, combined with the point estimate of the tariff effect for non-union workers, imply that relative to mining industries, which had tariffs close to zero in 1988, the wages in clothing declined by around 12% in clothing and 9% in furniture. How important are these kinds of effects in the context of the total CUSFTA package? Based on employment shares, the median tariff in 1988 was 4%. This figure, combined with the point estimate of the tariff coefficient, indicates that the median worker experienced a relative wage decline of around 2.8% as a result of CUSFTA. Only 10% of workers were employed 31 in industries with tariffs in excess of 8.9%. For these workers, relative wage losses attributable to CUSFTA would be on the order of 6.3%. The CUSFTA reductions thus had a large effect on the relative wages of workers in the most impacted industries. However, as the tariffs were low for the majority of industries and workers, the overall impact of CUSFTA on the wage distribution in the goods sector was small. The discussion in the previous paragraph indicates that while there were big losers from CUSFTA, the overall wage impacts were modest. Gaston and Trefler (1997) and Beaulieu (2000) reach a similar conclusion about CUSFTA, but for a very different reason. They find no direct impact. In contrast, I show that once compositional effects are controlled for, there was a substantial direct impact. How-ever, high tariff sectors accounted for only a small portion of overall employment in tradeables, and as a result the overall impact of CUSFTA was quite small. These results still suggest that even when trade occurs between countries with similar labour market institutions and endowments, liberalization can have a big impact on the wages of impacted workers. For Canada, future proposals for trade pacts include countries with abundant unskilled labour and lax labour laws. Whether these conditions will result in even bigger impacts of tariff reduction remains an issue for future research. 32 o Avg. Can. Tariff Avg. Amer. Tariff 7.5 CD -4—» CO 01 i ro CD cn CD CD 3 A 2.5 H oH 1980 1985 1990 Year 1995 Figure 2.1: Average Tariff Rates, 1981-1998 Closing co I-c CO o CD E < Canadian Tariff Figure 2.2: The 1988 Tariff Structure 33 • - , I ! 1980 1985 1990 1995 Year Figure 2.3: The Gender Differential 34 Table 2.1: Average wages and tariffs, no individual controls Industry 1981 wage premium W* 1981 Can. tariff r AW* 81-88 A T 81-88 AW* 88-98 A T 88-98 Clothing 2 33 17.73 -0.01 -0.52 -0.07 -17.20 Textiles 2 46 10.73 0.05 -0.82 -0.04 -9.90 Furniture & fixtures 2 52 16.73 0.01 -4.13 -0.09 -12.60 Food &; beverages 2 56 4.20 0.06 0.00 -0.03 -4.20 Miscellaneous manuf. 2 56 7.60 0.11 -2.34 0.02 -5.26 Rubber k, Plastics 2 57 12.73 0.18 -3.83 -0.13 -8.90 Printing 2 59 4.63 0.09 -3.23 -0.08 -1.40 Metal fabricating 2 69 10.25 0.06 -3.45 -0.07 -6.80 Electrical equipment 2 69 10.38 0.04 -4.28 0.11 -6.10 Non-metal manufact. 2 69 5.73 0.11 -2.33 -0.07 -3.40 Wood 2 73 4.50 0.00 -1.80 -0.09 -2.70 Machinery 2 82 7.78 0.00 -3.08 -0.01 -4.70 Paper .2 83 7.23 0.16 -3.23 -0.01 -4.00 Transportation equip. 2 83 2.83 0.03 -0.53 0.04 -2.30 Chemicals 2 84 7.40 0.06 -1.80 -0.05 -5.60 Non-metal mining 2 87 0.65 -0.02 -0.15 -0.05 -0.50 Primary metal manuf. 2 89 5.58 0.02 -1.58 0.04 -4.00 Metal fuels mining 2 93 0.70 0.16 -0.30 -0.05 -0.40 Metal mining 2 96 0.18 0.09 -0.08 0.06 -0.10 Source: Author's own data. The mean log hourly wage is constructed using the sampling weights of the data sets. 35 Table 2.2: Worker demographics, "tradables" sector only Variable 1981 1988 1998 Union coverage 47.33 45.38 35.87 Male 72.71 73.34 72.21 Male Covered 52.39 50.10 40.25 0-8 19.03 13.12 6.27 Some or completed H.S. 55.57 53.42 39.64 Some Post-secondary 7.95 9.48 7.31 Post-secondary certificate 9.52 15.33 34.80 University Degree 7.93 8.64 11.98 16-19 4.78 3.69 2.89 20-24 16.08 11.22 7.95 25-34 28.66 32.17 25.97 35-44 20.94 26.16 32.85 45-54 16.81 16.78 21.44 55-64 12.17 9.67 8.42 65-70 0.56 0.31 0.49 Female Covered 33.86 32.42 24.51 0-8 21.45 15.34 9.02 Some or completed H.S. 58.70 53.09 43.76 Some Post-secondary 6.91 8.25 8.33 Post-secondary certificate 9.84 14.61 28.64 University Degree 3.11 8.70 10.26 16-19 7.50 3.25 1.89 20-24 17.80 13.10 6.75 25-34 27.48 34.38 28.54 35-44 23.00 26.00 32.87 45-54 15.30 18.01 21.72 55-64 8.18 5.04 8.04 65-70 0.75 0.22 0.19 Source: Based on author's data. 36 Table 2.3: Union coverage and gender breakdown by industry, selected years Industry 1988 Can. tariff Coverage rate Male 1981 1988 1998 1981 1988 1998 Clothing 17.20 0 43 0 51 0 28 0 22 0 26 0 25 Furniture & fixtures 12.60 0 31 0 27 0 23 0 77 0 75 0 78 Textiles 9.90 0 46 0 45 0 38 0 46 0 59 0 59 Rubber & Plastics 8.90 0 37 0 40 0 30 0 71 0 79 0 73 Metal fabricating 6.80 0 44 0 46 0 28 0 87 0 87 0 84 Electrical equipment 6.10 0 44 0 34 0 26 0 61 0 70 0 68 Chemicals 5.60 0 30 0 30 0 18 0 73 0 66 0 59 Miscellaneous manuf. 5.26 0 37 0 26 0 19 0 59 0 60 0 61 Machinery 4.70 0 40 0 31 0 27 0 80 0 82 0 87 Food & beverages 4.20 0 41 0 46 0 38 0 64 0 66 0 62 Paper 4.00 0 70 0 73 0 61 0 87 0 89 0 85 Primary metal manuf. 4.00 0 58 0 54 0 54 0 94 0 83 0 89 Non-metal manufact. 3.40 0 44 0 55 0 42 0 82 0 82 0 89 Wood 2.70 0 52 0 48 0 34 0 91 0 89 0 88 Transportation equip. 2.30 0 67 0 62 0 56 0 88 0 81 0 80 Printing 1.40 0 37 0 26 0 22 0 56 0 51 0 59 Non-metal mining 0.50 0 67 0 49 0 37 0 93 0 85 0 92 Metal fuels mining 0.40 0 36 0 23 0 21 0 74 0 80 0 73 Metal mining 0.10 0 67 0 60 0 55 0 92 0 93 0 88 High tariff 10.25 0 41 0 40 0 29 0 61 0 66 0 65 Medium tariff 4.45 0 46 0 45 0 37 0 77 0 75 0 76 Low tariff 1.23 0 54 0 45 0 37 0 82 0 80 0 80 Source: Author's own data. Where appropriate, the sampling weights of the data sets have been used. 37 Table 2.4: Proportion lected years of unskilled and young workers by industry, se-Industry 1988 Can. tariff Unskilled Young (16-34) 1981 1988 1998 1981 1988 1998 Clothing 17.20 0 92 0 90 0.73 0 48 0 41 0 27 Furniture & fixtures 12.60 0 84 0 78 0.62 0 49 0 63 0 46 Textiles 9.90 0 85 0 81 0.61 0 50 0 44 0 31 Rubber & Plastics 8.90 0 79 0 72 0.52 0 56 0 54 0 38 Metal fabricating 6.80 0 81 0 71 0.45 0 51 0 50 0 45 Electrical equipment 6.10 0 67 0 48 0.33 0 45 0 49 0 35 Chemicals 5.60 0 58 0 47 0.29 0 44 0 47 0 34 Miscellaneous manuf. 5.26 0 73 0 66 0.39 0 60 0 48 0 42 Machinery 4.70 0 64 0 57 0.37 0 43 0 51 0 39 Food & beverages 4.20 0 82 0 72 0.56 0 54 0 50 0 41 Paper 4.00 0 83 0 65 0.43 0 41 0 37 0 27 Primary metal manuf. 4.00 0 74 0 66 0.44 0 54 0 43 0 24 Non-metal manufact. 3.40 0 79 0 76 0.58 0 47 0 54 0 29 Wood 2.70 0 88 0 80 0.61 0 53 0 59 0 43 Transportation equip. 2.30 0 72 0 67 0.44 0 45 0 43 0 36 Printing 1.40 0 72 0 58 0.50 0 57 0 54 0 44 Non-metal mining 0.50 0 83 0 76 0.67 0 54 0 39 0 35 Metal fuels mining 0.40 0 58 0 44 0.33 0 57 0 49 0 34 Metal mining 0.10 0 69 0 70 0.36 0 52 0 40 0 23 High tariff 10.25 0 81 0 73 0.54 0 50 0 50 0 37 Medium tariff 4.45 0 73 0 64 0.44 0 49 0 47 0 34 Low tariff 1.23 0 74 0 66 0.48 0 53 0 47 0 36 Source: Author's own data. Where appropriate, the sampling weights of the data sets have been used. 38 Table 2.5: Wage impact of CUSFTA, less skilled, non-union sector Specification (1) (2) (3) (4) (5) (6) Canadian .007 .008 .008 .008 .007 .008 tariff (.004) (.002) (.002) (.002) (.002) (.002) American .001 - - - - -tariff (.004) American .046 .051 - .051 .055 .049 employment (.050) (.048) (.044) (.044) (.048) Year yes yes yes yes yes no effects . Weights no no no no yes no Corrected for no no no yes no no Heterosked. Notes: Standard errors are in parentheses. There are 190 observations (19 industries by 10 years) in the second-stage regression. Table 2.6: Wage impact of Canadian Tariff, by skill and union status Low Skill High Skill Non-union Union Non-union Union Canadian Tariff Preferred 008*** .004** .007** -.001 specification (.002) (.002) (.003) (.004) Maximum .008** .005** .009** .001 (.002) (.002) (.003) (.004) Minimum .007* -.003 (a) .001 (b) -.001 (.004) (.003) (.009) (.004) No. of Obs. 20565 21045 15100 8522 in 1st stage Notes: * Significant at .10 level. ** Significant at .05 level. *** Significant at .01 level. (a) Specification including U.S. tariff. Estimate of .010 with s.e. of .003. (b) Specification including U.S. tariff. Estimate of .009 with s.e. of .005. 39 Table 2.7: Wage impact of CUSFTA, by skill and union status I. Low skill, Not covered by collective bargain Inexperienced Experienced a. Job Tenure 1)06** W)** (.003) (.004) 8138 12427 b. Age .005* .010** (.003) (.004) 9998 10567 II. High skill, Not covered by collective bargain Inexperienced Experienced a. Job Tenure .010*** ~002 (.004) (.005) 8255 6255 b. Age .003 .009* (.004) (.005) 7415 7695 Notes: When experience is defined by age, inexperienced workers are 16-34 years old, while experienced workers are 35-64 years old. When experience is defined in terms of job tenure, inexperienced workers have fewer than five years of tenure the current job, while experienced workers have five or more years of tenure. Table 2.8: The Effects of Controlling for Worker Characteristics (1) (2) (3) Canadian .003 .006 .008 Tariff (.003) (.004) (.003) Characteristic No Yes Yes Controls Time Variant _ No Yes 40 Table 2.9: Contr ibut ion of Changes to Composit ion and Returns of Worker Characteristics Tariff Classification Low Medium High 1981-88 Male .040 .031 .066 Educ -.002 .001 .005 Age .020 .035 .026 Tenure .061 .049 .055 Province -.004 -.016 -.018 Industry -.13 -.08 -.11 Total -.011 .021 .021 1988-98 Male -.072 -.051 -.068 Educ .014 .012 .008 Age .028 .037 .064 Tenure .035 .043 .046 Province .005 .024 .028 Industry -.025 -.154 -.132 Total -.015 -.090 -.054 1981-98 Male -.032 -.02 -.002 Educ .012 .011 .013 Age .048 .07 .090 Tenure .095 .092 .102 Prov. .000 .007 .010 Industry -.151 -.23 -.24 Total -.027 -.070 -.033 41 Chapter 3 Productivity, Labour Market Outcomes and Trade Liberalization: A Microdata Analysis 3.1 Introduction Much of the expected gains from the Canada-U.S. Free Trade Agreement (CUS-FTA) were expected to come through the realization of returns to scale. Large firms would expand their operations to exploit increasing returns to scale, while small, less productive firms would be driven out of business by the increased competition induced by the agreement. There is some evidence that scale increased in manu-facturing in Canada as the tariff reductions of the agreement were being phased in. Head and Ries (1999) find that average output per plant rose by 34% from 1988 to 1994. During the same time, the number of firms decreased by 21%. In a related study, Trefler (2001) finds evidence of increasing labour productivity in manufacturing during the CUSFTA period. Attempts to find an empirical link between trade liberalization and productivity gains experienced by Canadian manufacturing firms in the early 1990's have been mixed. Using data from the Annual Survey of Manufactures, Head and Ries (1999) find little evidence that tariff reductions accounted for the increase in output per plant that they uncover in the data. In their framework, expansion of output per firm and the decline in the number of firms appears to occur within both those industries that were highly protected prior to the agreement and those industries that had very little tariff protection. Although a link is established between the 42 tariff reductions and the increase in output per firm, increases brought about by lowering the American tariff are largely offset by contractions caused by decreasing the Canadian tariff. In contrast, Trefler (2001) finds that worker productivity increased after CUSFTA and that a significant amount of the labour productivity increase can be attributed to tariff reductions. In his framework, there is a 17 % increase in labour productivity in the most impacted industries. Of this, roughly half can be attributed to the tariff decreases mandated under CUSFTA. Further analysis by Trefler suggests that the increases are the result of favorable firm turn-over and not a result of firms expanding operations to take advantage of returns to scale. In this paper, I use data from micro-level labour surveys to examine whether there were tariff-related changes in employment patterns across industries that might account for changes in productivity across industries. In particular, I look at three outcomes and the extent to which developments in these outcomes can be accounted for by the tariff reductions mandated by CUSFTA. Outcomes of interest are the intra-industry employment shares across firms of varying size, the inter-industry union density and the inter-industry skills-distribution. The first measure may be though of as a crude measure of scale. Did the removal of protection result in an increase in the share of workers employed by large firms? Previously re-search has used the Annual Survey of Manufactures (ASM), which undersampled small firms over the early 1990s.1 Using a labour data set provides an alterna-tive measure of potential scale effects. Given the opposition by organized labour in Canada to CUSFTA, the effect of CUSFTA on unionization is of interest in its own right. Union leaders argued that free trade would result in de-industrialization in Canada, as firms relocated production to the southern United States to take advantage of lower labour costs and business friendly legislation. There is also a lit-erature examining the relationship between unionization and worker productivity. Unionization is thus a potentially valuable outcome to examine when accounting for productivity growth in the 1990's. Finally, I use educational data to examine the skills content of labour by sector to examine whether inter-industry changes in outcomes can be attributed to tariff changes. Previous work examining productiv-ity has distinguished between production and non-production workers. These are imperfect measures of worker skills. One possibility is that productivity gains re-l T n principle, the ASM is a census of manufactures. However, as a result of budgetary con-straints, a number of greenfield entrants, which tend to be small, were not added into the ASM until the mid-90's. See Head and Ries (1999) for a discussion. 43 suited from an improvement in the skills of workers within these categories that are not adequately accounted for when using the categories available in establishment level data. The remainder of this chapter is organized as follows. In section 3.2 I review the relevant literature to outline the expected implications of CUSFTA on each of the outcomes analyzed in this paper. In section 3.3 , I describe the data that is used in the subsequent analysis. In section 3.4, I provide a brief overview of the methodology to be used in subsequent sections. In section 3.5, I then use the labour data to examine whether movements in the data are consistent with the predictions. Section 3.6 provides a summary of the results and concluding comments. 3.2 Theory 3.2.1 Scale Simulations of CUSFTA suggested that the agreement would result in an in-crease in Canadian GDP of around 3%.2 Most of these gains were based on in-creased productivity achieved as a result of plant rationalization and scale effects.3 This kind of mechanism suggests that after the agreement, there should have fewer plants, but that the remaining plants should have been operating at a larger scale. Measuring scale effects may in practice be quite difficult. The usual proxy that is used in the literature is some measure of size. Trefler (2001) proxies scale with data in which plants are broken down by employment size. In his paper, growth in labour productivity over the 1988 to 1995 period is decomposed into growth occurring within size cells and growth occurring as a result of market share shifting between plants in different size categories. He finds that most of the growth is a result of increasing productivity within size categories and not as a result of production being shifted to the largest plants. As employment is the proxy for scale and there is no evidence that market share shifted to high employment plants, Trefler concludes that scale effects were not important in accounting for productivity gains in the period after CUSFTA was ratified. Head and Ries (1999) use output per plant as the measure of scale. Drawing on a number of models of trade in the presence of imperfect competition and scale 2See Hazledine (1990) for a summary and critique of these studies 3 See Chapter 2 for a more detailed explanation 44 economies, Head and Ries examine the relationship between the size and scale of Canadian plants and the magnitude of the Canadian and American tariffs. The expected impact turns out to be sensitive to assumptions about (i) whether or not new plants can enter freely, (ii) the extent to which Canadian and American markets are segregated, and (iii) the degree of substitutability between goods sold by different producers. In most specifications, reducing the home (Canadian) tariff is expected to decrease the number and scale of domestic plants, while reducing the foreign (American) tariff is expected to have the opposite effect. These predic-tions are opposite to those made by the computational general equilibrium (CGE) models used to justify the agreement. The C G E models predicted that reduction of the American tariff would increase the scale of Canadian plants. Using data from the Annual Survey of Manufactures (ASM), Head and Ries find that output per plant increased dramatically after 1988. Furthermore, they find a significant relationship between the tariffs and plant scale. Head and Ries find that cutting Canadian (American) tariffs caused Canadian plants to decrease (increase) plant scale economies. This finding is at odds with the argument made by CUSFTA supporters that reducing Canadian tariffs would result in increased scale. In any case, CUSFTA does not appear to account for the observed increase in output in scale, as the effects of decreasing the Canadian and American tariffs were offsetting. This suggests that some other phenomenon, such as technological change, accounts for the changes in scale. The results of Head and Ries (1999) must be viewed with some caution, as Statistics Canada missed some new entrants in the early 1990s as a result of budgetary constraints.4 New entrants tend to be small, and their inclusion would increase the number of plants far more rapidly than it would increase output, thus leading to a decrease in output per plant. However, Head and Ries find that output per plant still increases markedly when limiting attention to plants that employ 20 or more employees. As many entrants are small and would be excluded by this criteria, this finding suggests that output per plant did actually increase after CUSFTA was ratified. Baldwin et al. (2001) offer an alternative measure of scale. Rather than looking at a measure of size, they consider the number of product lines, or commodities, that individual plants produce over time. Presumably producing fewer commodi-ties allows a plant to increase the quantity produced of the remaining commodities. 4 The A S M was originally created for the purposes of building the System of National Accounts and has only recently been opened up to researchers for alternative purposes. 45 This in turn may be interpreted as a scale effect. Baldwin et al. find that plants produced fewer product lines starting in the late 1980s. Plant-level specializa-tion was most pronounced in those plants that increased their export intensity. This finding is presented as evidence that trade liberalization resulted in increased specialization at the level of the plant. The existing evidence suggests that CUSFTA did not result in the kind of rationalization effects that were predicted by the general equilibrium models. The studies summarized above used the ASM, which under-sampled small plants over the early 1990s. I re-examine this issue using labour data.5 If plant rationalization induced by CUSFTA results in a shift towards greater scale, then the proportion of workers employed by large plants (defined here by employment) in impacted industries should have increased disproportionately relative to the proportion of workers employed by large firms in unaffected plants. Unfortunately, due to data limitations, I work with firm size rather than plant size. 3.2.2 Unionization None of the economic models of CUSFTA that I am familar with made pre-dictions about how the implementation of CUSFTA would impact union density in Canada. Opponents of CUSFTA argued that the agreement would result in relocation of production to the Southern States, where "right to work" legislation and other institutional arrangements make it difficult to unionize. Furthermore, organized labour was one of the leading interests opposing the agreement. This suggests that there was at least a belief that the union sector would be negatively impacted by the agreement. The desire for representation by a union is often modeled as being determined in a cost-benefit framework.6 Workers choose to unionize if the benefits of doing so exceeds the costs. Works by Brander and Spencer (1988) and Mezzetti and Dinopoulous (1991) suggest that trade liberalization may affect the benefits of union representation. The wage that the union can procure for its membership will likely be sensitive to changes in the product market of the firm. If trade liberalization leads to increased competition, then firm level rents may be reduced. To the extent that the union wage reflects rent-sharing, one might expect that 5More recent studies by Baggs (2002) and Gu et al. (2002) use an alternative data set based on tax records for firms. These papers also fail to find a relationship between tariffs and firm size. I describe these papers later in this chapter. 6e.g. Farber and Saks (1980). 46 union wages would fall as well. 7 As Mezzetti and Dinopoulous illustrate, trade liberalization may also create the possibility to relocate production, which the firm can exploit to discipline union wage demands. If trade liberalization negatively affects the ability of unions to obtain high wages for their members, then one might expect that tariff reductions would lead to a lower union density within affected industries. Furthermore, if workers recognize that a successful union drive will result in the firm relocating, they will not choose to organize in the first place. These models suggest that CUSFTA could have had a negative impact on the union density in Canada. The organization decision is likely more complicated and depends on a variety of other environmental features, including labour laws, systematic preferences over unionization by worker type, and unobserved workplace characteristics. Johnson (2000) notes that there have been a variety of legislative changes governing the certification process that might potentially affect union density within Canada in the last 20 years. For example, in some jurisdictions, a card check is sufficient for unionization, while in other jurisdictions, a mandatory vote is required before union representation is obtained. Martinello finds a decline in organizing activ-ity in Ontario that follows changes in the' provincial laws governing certification. Interestingly, he also finds a decline in activity that follows immediately the im-plementation of CUSFTA. 8 The effect of unionization on productivity is still debated. Freeman and Medoff (1984) that unions increase productivity, by reducing quit rates and improving worker morale. Others argues that union practices such as ' feather bedding". 9 result in production inefficiencies. Grout (1984) suggests that the presence of unions result in inefficient investment in physical capital. 1 0 It has been noted that worker turnover is lower within unionized establishments. This has led Freeman and J.Medoff (1984) and others to argue that employers are more willing to invest in training if their workforces are unionized, as the investment embodied in the worker is less likely to leave the firm. The relationship between unionization and productivity is not examined fur-7The relationship between wages and tariffs was considered in the previous chapter. 8Unfortunately the data on union activity is not sufficiently disaggregate at the industry level to provide for a more thorough investigation. 9 "Featherbedding" refers to arrangements in union contracts where the number of workers required to work a machine is specified. 1 0It should be noted that Trefler (2000) finds no evidence that the productivity gains were attributable to capital deepening. 47 ther in this paper. I limit my investigation to examining whether union density changed more quickly in industries that where heavily protected prior to CUSFTA. I consider this issue by comparing the probability of a worker being unionized be-tween those industries that were heavily impacted by CUSFTA and those that saw little change in the rate of protection. 3.2.3 Skills Content In the Heckscher-Ohlin framework, trade is based on differences in factor en-dowments. Trade will result in each country exporting those goods which make intensive use of the factors in which it has a relatively abundant endowment. In turn, each country imports goods which make intensive use of factors for which it has a relatively scarce bundle. The transistion from autarky to trade affects both the price of prouducts (the autarky price is replaced by the world price) and the price of factors (through domestic factor markets). Trade will tend to increase the relative price of abundant factors (which are relatively scare on global markets). Al l industries will respond to changes in the relative price of factors by substi-tuting away the more abundant (and now more expensive) factor. This results in production in each good using relatively less of the abundant factor. Venables and Markusen consider a model in which multinationals are able to break up prouduction into vertical chains that span across countries. If different production stages involve different factor intensities, and if transportation costs are relatively small, then it may be optimal to locate different stages of production in different countries to take advantage of arbitrage opportunties arising from in-ternational differences in factor prices. Differences could result from differences in relative factor endowments, as in the Heckscher-Ohlin model, or from institution features, such as differences in unionization11 or legislated minimum wages. Patterns of trade based on differences in factor prices are usually considered important when considering trade between countries at different stages of develop-ment, where the relative scarcity of capital and skilled versus unskilled labour are thought to be important determinants of the trade flow. These kinds of factors are less likely to be important when considered CUSFTA, where both parties to the agreement are highly-industrialized countries with well- educated labour forces. However, differential taxes on capital, along with differences in unionization, may have been sufficient to distort relative factors prices between the two countries. nSee, for example, Staiger(1989). 48 In the presence of such distortions, CUSFTA might have resulted in changes in the ratio of skilled to unskilled workers between countries. For example, Canada tends to have higher union densities. Unions tend to compress the wage distri-bution, with relative gains for those less skilled workers on the lower tail of the distribution. This implies that the relative wage for less skilled workers would be potentially less in the United States than in Canada. CUSFTA would then tend to shift production that is labour intensive into the United States. The remaining production in Canada in industries that underwent this shift would tend to be skill-intensive. A complete investigation into relative factor endowments between Canada and the United States is beyond the scope of this paper. I use the educational attain-ment of workers to determine whether there was a relationship between changes in the skills composition of industries and the change in the tariff structure. It is not, however, clear that this relationship should exist on an industry by industry basis; general equilibrium effects of trade liberalization could potentially result in changes in skills compositions across all industries. As will be seen, this kind of research is further complicated by a general increase in educational attainment in Canada. These trends indicate that in the absence of CUSFTA there would have been a general increase in the ratio of skilled to unskilled workers in most industries. 3.3 Data I make use of a collection of labour force surveys for Canada. These surveys consist of the Canadian Labour Force Survey (LFS) and a number of supplements to the LFS. In particular, I also make use of the 1986/87 Labour Market Activities Survey (LMAS), the 1988-90 LMAS, the 1993-94 Survey of Labour and Income Dynamics (SLID), the 1995 Survey of Workplace Activities (SMA), and the 1993, 1994, 1997 and 1998 LFS. These surveys are administered as supplementary sur-veys to the LFS; as a result they are similar in design and generally comparable. Not all variables are available in all data sets. For example, the 1993 and 1994 LFS do not include information on firm size or union status, while the 1993-94 SLID includes educational categories that are not directly comparable with those used in other survey years. By making use of multiple sources for 1993-94, the data sets employed here provide for a fairly comprehensive picture of the outcomes of 49 interest for a number of years spanning the period 1986-98. My data set consists of those individuals that were employed as paid workers at the time of the survey. Al l information applies to the main job of the individual, which is defined as the job with the most usual hours; To further insure compat-ibility across surveys, I use the employment status of individuals in the second week of November. Further details are provided in chapter 2. 3.4 Methodology In the work which follows, the investigation of a link between tariffs and a given outcome follows a standard "control-treatment" approach. To be precise, I am comparing the outcomes in industries which were previously protected but no longer are with the outcomes of those industries that have historically received little or no protection. There is some flexibility in choice of a "control" group. This group may be either limited to those industries that are exposed to international trade directly (the tradable goods sector) or expanded to include industries that produce services. My inclination is to opt for the former group, since the latter group is not exposed to competition through trade in the same manner as goods producing industries. There is clearly trade in services and a variety of laws which place restrictions on their trade. Without a clear way to measure these restrictions, I am forced to assign non-goods a tariff of zero throughout the sample period. This is a severe limitation which I feel merits the exclusion of the services industries from my analysis. Furthermore, much of the non-good producing goods sectors are in the government sector or are regulated by the government (e.g. utilities). The role of the government in these industries, along with recent pushes towards privatization and deregulation, suggest that these industries may have behaved much differently than those in the goods sector. The outcomes considering in this chapter are all reported categorically in the data set. In the case of unionization, this is natural; the conditions of a worker's job either are or aren't determined through collective bargaining. In the case of the size of the firm of employment, the categorical nature of the data is a feature of the data set. Statistics Canada codes a number of outcomes categorically to prevent use of the information to identify individual workers or employers. In the case of educational attainment, a similar rational holds. Given the categorical nature of the data, the link between a particular outcome 50 and the applicable tariffs is analyzed using either a simple probit or an ordered probit. Consider the union status of worker i in industry j at time t. In section 3.2.2, I reviewed an approach used in the industrial relations literature in which workers choose union representation based on a cost-benefit basis. Suppose that this decision is influenced by industry level conditions including available technol-ogy, which dictates the nature of the production floor, and available rents, which influence the extent to which a union can elevate the wage rate. In this case, one can consider a latent index for each worker, denoted by Zijt, which depends on industry conditions such as the tariff rate Tjt. Let this index be given by zijt = XjtB + Vjt + tit (3.1) Further assume that a worker will obtain union status if this index assumes some threshold value. In this context, the index can be interpreted as worker i's utility of unionization. Other outcomes can be handed in a similar fashion using an ordered probit. In particular, consider plant size. For each worker i in industry j at time t, consider an index of the form ZJU = Tjt3 + Ya + ta + iiju (3.2) where Tjt is a vector of Canadian and American tariffs for industry j in year t, Yt is the year-specific unemployment rate and t is a linear trend. fijit is assumed to be normally distributed, but is allowed to be correlated across all individuals in a particular industry-year cell. Let d denote the firm categories. Then the probability of an individual being employed at a firm of size d while employed in industry j at time t is given by P(firmsizejit = d) — P (K d _ i < zjit < Kd) (3.3) where «d are a series of cut-off points over the normal distribution to be estimated along with the parameters of the index equation Zjit. 51 3.5 C U S F T A and Selected Outcomes 3.5.1 Plant Scale Table 3.1 presents a breakdown of employment shares by firm size 1 2 in 1988 and 1998 for low, medium and high tariff industries, along with an aggregate category for all remaining industries in the economy.13 The years in question are chosen because they occur at similar points in the business cycle. For the first two industry groupings (low and medium tariff), employment shares by firm size appear fairly stable. In comparison, there appears to be a shift in employment shares in both the high tariff industries and the rest of the economy away from small firms (20 or fewer employees) towards very large firms (500 or more employees). This would be consistent with a reallocation towards larger firms, possibly to exploit economies of scale. As can be seen from Table 3.1, sectors that have historically faced international competition have a large share of employment with large firms, while firms that were previously protected by high tariffs have a greater concentration of employment among small- and medium-sized firms. There is a limitation to using firm size as a measure of scale, in that it is not possible to observe reallocations within the largest category. In particular, Head and Ries (1999) argue for the possibility of two types of firms within industries: small firms which serve local niche markets and large firms which produce for a mass international market. If this were the case, trade liberalization might only impact the largest firms, with workers being relocated from large firms to a handful of even larger firms. Such a relocation would not be detectable with the firm size categories used here. To examine more formally whether tariff reductions have resulted in employ-ment shifts towards larger firms, I make use of several discrete outcome estimators. First, I define an indicator variable for all workers with firms that employ 500 or more workers, which I consider to be a large firm.14 I use this variable to esti-mate a probit in which this outcome is conditioned on industry specific tariffs, a 1 2The ASM data is based on establishment size rather than firm size. Unfortunately this data is not available for most of the years of data that I have, due to Statistics Canada's policies on individual confidentiality. 13Low tariff industries are defined as those industries with a Canadian tariff below 3% in 1988, while high tariff industries are those industries with a Canadian tariff above 6% in 1988. 1 4In my research, I tried several alternative specifications. In particular, I defined an alternative indicator for all workers with firms that employ 100 or more workers and repeated the same analysis described above. The results were similar when using this alternative definition of a large firm. 52 linear trend, the economy-wide unemployment rate, and industry fixed effects. A specification is included in which the American tariff is omitted. As illustrated in the previous chapter, the Canadian and American tariffs are highly correlated, suggesting that identification of separate tariff effects may not be possible. A specification is estimated which precludes the American tariff. In this case, the Canadian tariff coefficient may be interpreted as the combined effect of reducing the Canadian tariff and making the offsetting reduction to the American tariff. Head and Ries (1999) found that currency depreciation accounted for much of the rise in output per firm. To allow for this, I consider an alternative specification which includes the exchange rate. I adjust the standard errors of the estimates to allow for a common industry specific shock in each year. The resulting estimates of the specifications described above are presented in Table 3.2. Specifications 1 and 2 are estimated without the exchange rate. Specifi-cations 3 and 4 include the exchange rate. Tariffs are not a significant determinant of employment in any specification. The Canadian tariff coefficient is negative, indicating that higher tariffs are correlated with more employment among smaller firms. Clearly this is consistent with the notion of trade liberalization resulting in the large scale operations. However, in no specification are the tariff coefficients significant. Neither the unemployment rate nor the exchange rate are significant predictors of firm size employment shares. As suggested by Table 3.1 there is not a great deal of evidence of a shift of employment towards larger firms over the study period; it is thus not surprising that the trend is also not significant. I re-examine the effect of tariff reductions using the full range of responses. As the categories are ordered, I use an ordered probit to investigate any relationship. The parameter estimates are reported in table 3.3 for four specifications. As be-fore columns 1 and 2 are estimated without the exchange rate, while columns 3 and 4 include it. Columns 1 and 3 are estimated including the American tariff, while columns 2 and 4 are estimated without the American tariff. As expected, the tariff coefficient is negative, suggesting that higher Canadian tariffs increase employment shares for smaller firms. However, the coefficient estimates are once more statistically insignificant. Furthermore, there is little evidence of an overall shift towards larger firms. On the whole, my findings are consistent with those of previous researchers. The available evidence suggests that the CUSFTA reductions had no impact on scale, as measured by the number of employees at either the level of the establish-53 ment or the firm. These results suggest that the productivity growth measured by Trelfer (2000) resulted from some other mechanism. 3.5.2 Unionization Union density in Canada began to decline in the 1990's, after exhibiting sta-bility throughout the 1980's. Figure 3.1 shows the year by year union density for the same industry categories used in the previous section. For all sectors, union density is reasonably stable throughout the late 1980's. There is evidence of a decline occuring in most sectors throughout the 1990's.15 The high tariff sector is less heavily unionized than the rest of the tradeable goods industries throughout the sample period. There is some indication of a more pronounced decline in union density in these industry grouping as well. A detailed model of union status is beyond the scope of this paper. My intent is simply to determine whether or not there is a correlation between the inter-industry union density and the tariff structure. As Farber and Saks(1980)have noted, the desire for union representation varies systematically across observable characteristics. For example, workers on the low end of the wage distribution may view unionization as more advantageous than their better paid counterparts. To deal with heterogeneity of worker characteristics across industries, I consider a simple model in which union status depends on industry characteristics and individual characteristics. Consider an index U*JU, Uju* = Xit8 + ZjtX + Yta + tjit (3.4) where Xu is a vector of individual characteristics, such as gender and education, Zjt is a vector of industry specific characteristics, including tariffs, and Yt are common time- varying variables, is assumed to be distributed normally. As before, the error is allowed to be correlated across individuals within industry-year cells. An individual is unionized if his or her index is greater than zero. The residuals are assumed to follow a normal distribution, leading to a standard probit specification. Table 3.4 presents results for the a probit specification in which individual union status16 depends on the characteristics of the industry of employment and possibly individual characteristics. Industry specific fixed effects are included. The first two columns are estimated excluding individual characteristics. The last 15Johnson(2001) also finds evidence the Canadian union density began to decline in the 1990's. 1 6I use coverage by a collective bargaining agreement as my measure of unionization. 54 two are estimated including information on an individual's gender, province of residence, educational attainment and age. Two points are worth noting. First, tariffs do not have a significant impact on union status. Second, there is a common trend towards de-unionization. To establish the extent of this trend, I first predict the industry average union density for 1998, using specification 3. I then predict a value for union density replacing the trend variable with its 1988 value. The first value is 30%, while the second value is 35%. This suggests that there was considerable de-unionization in manufacturing that cannot be accounted for by either the CUSFTA tariff reductions or shifts in workforce composition. In chapter 2, I proposed that the higher tariff impact on wages in the non-union sector could be accounted for if unions were responding to decreased labour demand by trading employment for wage maintenance. If this were the case, one might expect to see the union densities fall in those sectors that faced the largest tariff impacts. The results presented above provide no support for this hypothesis. Instead, there is apparently some kind of cross-industry phenomena that is driving this decline. 3.5 .3 Skills allocation The production/non-production worker categories are only imperfect proxies for skills categories of workers. Increases in worker productivity may be related to increases in average skill levels within production and non-production worker groups. This effect will not be captured if the input of production workers is not quality adjusted through some kind of hedonic wage measure. As the LFS contains detailed information on educational attainment of workers, I use this variable instead as a measure of worker skill. The educational variable recorded on the LFS may be broken into five ordered categories: 1)0-8 years of schooling, 2) 9-13 years of school and/or completed high school, 3) some-post secondary education but no degree, 4) post-secondary certificate, and 5) university degree. In 1990 the LFS educational question was redesigned in a manner that resulted in some changes in response patterns.17 Prior to 1990, respondents were only asked about post-secondary education if they had completed high school. After the change, all respondents were asked about post-secondary training. Several kinds of reclassification were possible between the two survey questions. Potentially, the most important changes involved individuals without a high-school diploma but 17See chapter 2 for more details. 55 with some form of training that was considered post-secondary for the purposes of the survey. These individuals would have been in category 2 in the old scheme but in category 4 in the new scheme. Table 3.5 provides tabulations on workforce distribution of skill for 1988 and 1998 for the four industry groups used previously. As should be immediately obvious, there appears to be a common tendency towards a more skilled workforce from 1988 to 1998. The growth in the fourth educational category and the decline in the second category is particularly striking, and is consistent with the change in the survey design. However, there is little evidence of a systematic difference in this shift across industry grouping. The most protected industries tend to heavily utilize workers with little or no education.18 I investigate educational shares in the same manner as firm size. In particular, an ordered probit is used in which the outcome variable consists of the five educa-tional categories described above. To account for the redefinition, a indicator value is included which takes on a value of '0' for the years 1986-90 and a value of '1' for the years 1993-1998.19 The results of estimating this model are shown in Table 3.6. The first specification includes both tariffs. The second specification is similar, and also includes the indicator for those years after 1990. Columns three and four are analogous to columns one and two, but exclude the American tariff. Before discussing the tariff coefficients, it is worth noting that the indicator variable for the split period is not statistically significant and that including it has little affect on other variables, including the trend. This suggests that the movement across educational groups between 1988 and 1998 is the result of an actual increase in educational attainment over time, and not simply an artifact of survey redesign. As a further test, I use the second specification to predict the 1998 educational distribution for the high tariff sector by first using the actual values of the covari-ates and then by setting the indicator to the pre-1980 value. Results are shown in Table 3.7. The predictions are quite similar. Furthermore, using the pre-1990 indicator results in a higher level of educational attainment, which seems to run counter to the effects that one would expect, given the nature of the redesign. 18There is nevertheless a fair bit of heterogeneity within groups. The electrical equipment industry is included in the high tariff grouping. In 1988, 17% of workers in this industry had a university degree and 42% had at least a post secondary certificate. This is in sharp contrast to the clothing industry, where only 2% had a university degree and only 7% had at least a post-secondary education. 1 9The new educational question was introduced into the LFS in 1990. The 1990 LMAS edu-cational question is based on the pre-1990 design. 56 The tariff co-efficients are generally not significant at generally accepted levels of significance. The only exception is the American tariff, which is positive and significant at the 10% level of significance and has a positive coefficient. Taken at face value, this would suggest that' reducing the American tariff reduces the skill content of the affected industry. Keeping in mind that the Canadian and American tariffs were reduced simultaneously, I examine the implications of these estimate as follows: 1) I predict the skills content in the high tariff industry for 1988 using the actual tariff values. This corresponds to checking how well the model predicts the actual values observed in 1988. Results are presented in row 1 of Table 3.8. Comparing this for the results in table 3.5 indicates that the model provides a reasonable fit. 2) I predict the skills content using the 1988 Canadian tariff and an American tariff rate reduced by 1 basis point below the 1988 value. This corresponds to evaluating the marginal effect of a 1 basis point reduction in the American tariff. The result is shown in row 2 of Table 3.8. Comparing this result with the first row suggests that a one basis point reduction has a very tiny effect on the distribution; workers are now more likely to be employed from the lower end of the distribution. 3) I predict the skills content using tariffs 1 basis point below the 1988 value for both tariffs. Results are shown in row 3 of the same table. When compared to the second row, this shows the marginal effect of reducing the Canadian tariff. Again the effect is tiny; a comparison with row 1 suggests that the reduction in the Canadian tariff cancels out the effect of reducing the American tariff. 4) I set the American tariff to zero but hold the Canadian tariff at its 1988 level. The results are shown in row 4. By comparing this result with row 1, one can see the cummulative effect of phasing out the American tariff. Again, the results are small, but again phasing out the American tariff reduces the skill content in the most impacted industries. 5) I set both tariffs to zero. The results of this exercise are shown in row 5. When compared to row 1, this shows the cummulative impact of CUSFTA on the most impacted industries. As a comparison with row 1 indicates, CUSFTA appears to exert a negligible net effect on the skills distribution. On the whole, the results presented here suggest that CUSFTA had no impact on the skills distribution in impacted sectors. This contrasts somewhat with Trefler (2000) and Beaulieu (2000), who both found evidence of relative skill upgrading in the impacted sector. My results are not neccessarily at odds with those of the previous papers, as I use an alternative measure of skill, based on a worker's 57 educational attainment rather than his of her job classification. However, for the two results to be consistent, it would have to be the case that production workers were becoming more educated over time in these industries. These results suggest that production/non-production worker categories may not be good proxies for worker skill levels over longer time horizons. 3.6 Discussion and Conclusion CUSFTA was expected to result in significant productive gains as increased competition drove small firms out of business and forced large firms to increase their operations to realize returns to scale. Using data from firms, previous researchers found little evidence that tariff reductions resulted in increases in scale in the affected industry. In this paper, I contribute to this research by re-examining the effect of CUSFTA on scale. I also examine the impact that CUSFTA had on unionization and skills mix in affected industries. My general finding is that CUSFTA had no impact on any of the outcomes under consideration. Possible scale effects were investigated by looking at the employment patterns across industries by firm size. I find no evidence that the importance of large employers increased in those industries most impacted by CUSFTA, which is at odds with models by Cox and Harris (1984) and others that predicted large scale effects would result from the agreement. A growing number of studies, using a variety of sources of data, have been unable to produce any evidence that CUSFTA led to substantial increases in scale economies. These studies include Head and Ries (1999) and Trefler (2001), which used data from the A S M . Recent research by Baggs (2002) and Gu et al. (2002) have also examined the issue using data compiled from tax records and have also been unable to establish a relationship between firm size and tariffs. My findings, obtained using a completely different source of data, provide further evidence that CUSFTA did not have the effect on scale economies that early models suggested. Rates of unionization appears to be unaffected by CUSFTA. In principle, CUS-FTA could have reduced the payoff to workers of organizing in the most impacted industries. However, it appears that union density was falling across all industries, regardless of the degree to which tariff protection was cut. This suggests that whatever forces were at work in reducing the rate of union representation were not related to CUSFTA. Possible candidates for de-unionization include key changes 58 in the legislation that governs union certification. The skills mix of employees also appears to be unaffected by the CUSFTA cuts. Educational attainment increased in all industries. An empirical framework provides no evidence of a systematic relationship between this upgrading and the tariff cuts. This results are seemingly at odds with previous research, which used production and non-production workers as proxies for unskilled and skilled work-ers respectively, and which found accelerating skills upgrading in those industries where the tariffs were most heavily cut. These results would be consistent with mine if production workers were gradually becoming more educated over time. On the whole, CUSFTA appears to have had a minor impact on the outcomes in individual industries of the Canadian economy. However, I consider outcomes that are measured in terms of ratios. CUSFTA may have resulted in substantial reallocation of resources across industries. Furthermore, the most important as-pects of CUSFTA may not be easily quantified. For example, it has often been argued that the most important feature of CUSFTA was that it provided Canada with guaranteed access to the American market by providing a mechanism for con-flict resolution. In this case, the impacts may not be related to industry-specific measures of the tariff protection. As noted, productivity gains occurred across Canadian manufacturing indus-tries after CUSFTA was implemented. Two potential explanations, examined here, are de-unionization and increased skills among workers. I find evidence that both of these phenomena were happening during the study period. However, there is little to suggest that these phenomena were confined to the industries that were directly affected by CUSFTA, or that tariff reductions associated with CUSFTA in any way expedited these developments. Whether either of these phenomena can account for productivity gains over the CUSFTA period remains a subject for future research. 59 60 Table 3.1: Employment share by firm size, selected industries and years Sector Year 1-20 20-99 100-500 500+ Low Tariff 1988 12.4 17.3 16.6 53.7 1998 13.2 15.7 18.3 52.7 Medium Tariff 1988 9 19.5 21.4 50.1 1998 10.3 18.5 20.0 51.2 High Tariff 1988 17.1 27.4 21.9 33.7 1998 13.2 26.5 22.2 38.2 Rest of economy 1988 37.2 16.9 11.2 34.8 1998 30.4 17.4 13.5 38.7 60 Table 3.2: Probit analysis of likelikhood of being employed by a large firm 1 2 3 4 Canadian tariff -0.014 (.013) -0.011 (.008) -0.014 (.013) -0.011 (.008) American tariff .00 (.01) - .00 (.01) -Unempl. rate .00 (.01) .00 (.01) .00 (.01) .00 (.01) Exchange rate - - .00 (.13) .00 (.13) Trend -0.003 (0.005) -0.003 (.005) -0.003 (.006) -0.003 -0.005 Pseudo R2 0.12 0.12 0.12 0.12 Notes: *Denotes 5% level of signficance. 61 Table 3.3: Employment share by firm size and tariffs 1 2 3 4 Canadian tariff -0.005 -.004 -0.005 -0.004 (0.013) (.007) (0.013) (.007) American tariff .00 .00 (.01) (.01) Unempl. rate .00 .00 .00 0 (.01) (.01) (.01) (.01) Exchange rate .00 .00 (.11) (.12) Trend .000 .000 .000 .000 (.005) (.005) (.005) (.005) Pseudo R2 0.07 0.07 0.07 0.07 n 44267 44267 44267 44267 J 19 19 19 19 Table 3.4: Effect of tariffs on union status 1 2 3 4 Canadian tariff .00 (.01) 0.007 (.007) .00 (.01) 0.01 (0.007) American tariff 0.01 (.01) 0.01 (.01) Unempl. rate 0.007 (.010) 0.07 (.010) 0.01 (.01) 0.01 (.01) Trend -0.022* (.003) -0.022* (.003) -0.018* (.004) -0.017* (.004) Individual Characteristics no no yes yes Pseudo R2 n J 0.13 52733 19 0.07 52733 19 0.07 52733 19 0.07 52733 19 *Denotes signficance at 1% level. 62 Table 3 . 5 : Skills distribution by industry 0-8 years 9-13, Some post- Post-sec. University secondary diploma Degree I. Low Tariff 1988 12.7 53.3 8.4 14.8 10.8 1998 5.3 43.1 7.2 34.5 9.9 II. Medium Tariff 1988 11.7 52.6 10.3 16.2 9.2 1998 6.1 37.2 7 35.9 13.8 III. High Tariff 1988 18.1 55.5 7.9 12.1 6.4 1998 11.8 43.2 6.6 28 10.4 IV. Non-tradables 1988 12 50 9.8 15.7 12.5 1998 4.7 38.2 9.6 31.9 15.6 Table 3 . 6 : Effect of tariffs on skills distribution 1 2 3 4 Canadian Tariff -0.008 (.007) -0.008 (.007) .000 (.005) 0 (.005) American Tariff 0.013* (.007) 0.013* (.007) - -Unempl. Rate 0.009 (.007) 0.03 (.02) .01 (.01) 0.03 (.02) Trend 0.04 *** (.004) 0.05*** (.01) 0 04*** (.004) 0.05*** (.01) After 1990 — -0.11 (.10) — -0.10 (.10) Notes: *Indicates significant at 10 % level of significance. **Indicates significant at 5 % level of significance. •^Indicates significant at 1 % level of significance. 63 Table 3.7: Simulated skills distribution, high tariff industries, 1998 0-1 I years 9-13 year, High-school Some post-secondary Post-sec. diploma University Degree Predicted 9.5 46.0 8.3 24.9 11.3 Pre-90 7.9 43.7 8.5 26.6 13.3 indicator Table 3.8: Simulated effect of tariffs on skill distribution, high tariff industries, 1988 Step 0-c I years 9-13 years, Some post- Post-sec. University Comp. H.S. secondary diploma Degree 1 18.0 52.4 7.0 17.2 5.4 2 18.2 52.5 6.9 17.0 5.3 3 18.1 52.4 7.0 17.2 5.4 4 20.4 52.4 6.6 15.9 4.8 5 18.1 52.3 7.0 17.2 5.4 64 Chapter 4 The Sources of Declining Entry Wages for the Less Educated in Canada With David Green 4.1 Introduction Over the past two decades, Canada has experienced a considerable increase in male earnings inequality. It is by no means alone in this trend, with the U K and the US experiencing stronger trends. For Canada, several papers have demonstrated that increases in cross-sectional age or experience earnings differentials over time have played a particularly large role in the increases in overall inequality. For example, Kapsalis et al. (1999) shows that the difference in average weekly earn-ings between 25 to 34 year olds and 45 to 54 year olds increased from 18% to 30% between 1981 and 1988. This has been interpreted as reflecting a substan-tial increase in the returns to the human capital for which experience acts as a proxy. However, Beaudry and Green (2000) demonstrate that the data fits with a specification in which successive birth cohorts face parallel but successively lower age-earnings profiles. The fact that these profiles are parallel, that is that they do not get steeper over time, means that no group of workers actually experienced in-creasing returns to age, contradicting the standard interpretation of age-earnings movements. Rather than reflecting increasing returns to skill, the age-earnings movements stem from continuously worsening earnings outcomes for successive cohorts of labour market entrants. This result is echoed in the U K ((Gosling et a l , 2000)) and the US (MaCurdy and Mroz (1995)). Thus, understanding the 65 changing experiences of more recent labour market entrants is a key component in understanding inequality movements in several countries. The finding that successive cohorts face similar age-earnings profiles but at lower levels suggests a model incorporating substantial labour market rigidities. With access to data-sets including continuous years of job tenure variables, we organize the data on the basis of job entry cohorts (cohorts of workers defined according to when their job started) rather than birth cohorts. We find that the data can be characterized by a simple specification with constant, across years, age-earnings and tenure-earnings profiles combined with declining average wage levels for successive job entry cohorts. In particular, we find that new job entrants in 1998 earned real hourly wages that were over 20% lower than those earned by new job entrants in 1981, regardless of their age at the start of the job. However, once on the job, workers' wages increased at the same rate with job tenure regardless of the starting date of the job. Again, this does not fit with returns to experience and job tenure related skills increasing over time. We argue that these patterns fit with a model in which a declining real wage is distributed among generations of new job entrants through fixed wage profile contracts set at successively lower levels for each successive cohort. In this paper, we also attempt to uncover what is behind this finding that successive job entry cohorts face substantially lower entry wages. As described by Morissette (1997a) in an investigation of outcomes across birth cohorts, and as confirmed here for job entry cohorts, declines in real wages across cohorts have been accompanied by falling rates of unionization and access to jobs in high paying industries. We seek to document these trends by job entry cohort and relate them to the declines in real wages. In doing this, we are interested not only in movements in mean wages but also in changes in the shape and position of the entire wage distribution. This is necessary to achieve a full accounting because of well-known effects of unionization on the shape as well as the location of the wage distribution (Lemieux (1998)). Focussing only on movements in the central tendency of the distribution may under-state the effects of shifts in unionization across cohorts. To carry out such an investigation, we require an estimator that permits estimation of covariate conditional cumulative distribution functions. We use a hazard based estimation approach developed in Donald et al. (2000). One advantage of this approach is that it provides a natural means for estimating the impact of minimum wages on the wage distribution. As we will see, minimum 66 wage impacts play a crucial role in interpreting cross-cohort movements in the wage distribution. We carry out our investigation using data on workers with a high school or less education gathered from a series of Canadian micro data-sets spanning the period from 1981 to 1998. The data-sets are chosen because they are very similar in sample frame and include information on union status, job tenure measured in years, and hourly wages. We focus on less educated individuals because this is the group for whom the observed changes in the cross-sectional experience differential and cross-cohort earnings declines have been the largest. We examine outcomes for males and females separately. Much of the existing literature focusses solely on male wages, however to obtain a full understanding of recent wage movements we need to incorporate female outcomes as well. The paper proceeds in six sections. In section 4.2, we describe results in earlier papers. In section 4.3, we describe the data. In section 4.4, we present several linear regression specifications for modelling the wage process and provide the results of specification tests that indicate that the representation described above performs well in capturing the major developments in the mean movements in wage distri-bution. In this section, we also document movements in covariate values across cohorts. In section 4.5, we provide a brief description of the covariate conditional distribution estimator we use. In section 4.6, we present results, showing first the marginal effects of key covariates on the wage distribution and then providing de-compositions of the distribution of wages for new job entry cohorts from the early 1980s and the late 1990s. Section 4.7 contains a summary and conclusions. The key findings from our investigation are as follows: 1) For successive male job entry cohorts, employment has shifted from the primary sector to services. Unionization rates for these cohorts have fallen as well, even within sectors, sug-gesting that union job openings are less and less available to more recent job entry cohorts, even after accounting for employment shifts. 2) About half of the decline in the real wage for successive job entry cohorts over the study period can be accounted for by declining unionization and changes in the industrial pattern of employment across job entry cohorts. A similar figure holds for both males and females, though there are differences in the pattern of employment between the sexes. In accounting for the effects of unions, we consider both the direct effect of unionization on the wages of members and spill-over effects of industry union density on the wages of non-union workers in the same industry. 3) The decline 67 in earnings among female job start cohorts appears to have been mitigated by the minimum wage, which effectively "shores up" the lower tail of the distribution. The minimum wage has had little effect on the change in male cohort earnings. Males tend to earn more than females on average and as a result most of the mass of the male wage distribution lies well above the minimum wage. 4.2 Previous Literature 4.2.1 An Empirical Framework We begin with specifying a simple empirical framework which will aid in the discussion of previous papers and the results in this paper. Consider a reduced form process in which wage movements over time can be conceptually decomposed into three components: cohort specific effects, age effects and time effects. The first of these refers to average, lifetime wage levels that are specific to a given birth cohort and might arise from factors such as differences in schooling quality across generations. The second component refers to lifecycle wage profiles that may reflect factors such as investment in and returns to human capital. The third component refers to macro-economic shocks that affect all workers, regardless of their cohort and age, in a given year.1 As is well known, one cannot work with a specification containing flexible representations of all three components because of perfect collinearity arising from the fact that the three components are related through an identity (once one knows any two of the current year, an individual's current age, and the individual's birth year (cohort) then the third is exactly determined). In addition, in our case we also wish to consider job tenure related effects. These potentially can be described in an analogous manner to age, using tenure on the current job, the job entry cohort (captured by the year the job started), and a year effect. Again, we can only include at most two of these effects in any empirical model. We will denote the real log wage of an individual as w, the current year as yr, their age in a specific calendar year as a, their birth cohort as b, their number of years of job tenure in the calendar year as t, and their job entry cohort as c. The relationship considered in previous papers can be summarized as, 1Note that these macro effects could take the form of either cyclical effects or longer term trends. All three papers being discussed here (i.e., MaCurdy and Mroz (1995), Beaudry and Green (2000) and Gosling et al. (2000)) use specifications that filter out cyclical effects in order to focus on long term trends. 68 w = f(a,b,yr) (4.1) It is perhaps easiest to consider this relationship in a mean regression context and think of the right hand side variables as being broken into discrete categories. In that case, if the / function represents all possible interactions of the discrete variables then / summarizes the wage-age-cohort relationship completely. The goal in most papers is to place meaningful restrictions on / in order to relate the data patterns to economic theory. Consider, first, a simple specification in which the log wage is generated from a process with each birth cohort facing the same wage-age profile but having its own intercept for that profile. where, w equals the log wage of individual i who is born into cohort, b, and is of current age, a. The / and g functions are flexible functions of age and a birth cohort index, respectively, and need not have positive first derivatives. Thus, this specification would allow for successive cohorts facing the same wage-age profile but having declining or increasing intercepts for that profile. Alternatively, one could make the specification more flexible by allowing the wage-age profiles to differ across cohorts. where, h(.,.) is a function reflecting differences in the wage-age profile across co-horts. Introducing an increase in the returns to experience in a given year, y, for a given cohort, b, would generate an increase in the slope of the wage-age profile at age (y — b) for this cohort. Other cohorts would experience the same increased slope but at different ages - the ages they were in year y. This would be reflected in the function, h(.,.). For an ongoing increase in the returns to experience, the derivative of h(.,.) with respect to a will be increasing in b for a given a. In the work that follows, we wish to examine implications for job entry cohorts as well as birth cohorts. We can introduce job entry considerations by adjusting specification (4.3) as follows, w = f{a)+g(b) (4.2) w = /(a) + g(b) + h{a, b) (4.3) w = f*{a) + g*{b) + h*{a, b) + A(c) + 7(1) + 5(c, t) (4.4) 69 where the first three terms on the right hand side of the equation have the same interpretation as in (4.3) but have been rewritten with an asterix to emphasize that results obtained from specification (4.3) may reflect unaccounted for tenure effects through correlations between a, b and tenure. In specification (4.4), c indexes year of job entry (i.e., job entry cohort), t equals years of job tenure as of the current year, and A(.), 7(.) and 5(.) are functions. Note that we cannot allow completely flexible forms for the functions on the right hand side of (4.4) because both a and b and c and t will add up to the current year. This specification can be identified in practice to the extent there are restrictions on these functions such as flat regions of the wage-age profile. Nonetheless, equation (4.4) provides a useful starting point for considering issues of movements in wages by age over time. 4.2.2 Previous findings A large number of papers establish the fact that the wage differential between more and less experienced workers has increased in the US in the last 20 years. For example, Juhn et al. (1993) find that the difference between wages for workers with 1 to 10 years of experience and workers with 21-30 years of experience increases by 20 percent between 1964 and 1988. Buchinsky (1994) uses quantile regressions to provide a complete representation of the experience differential across the earnings distribution. Both Buchinsky (1994) and other papers show that the increase in the experience differential is much larger for less educated workers. Schmitt (1995) finds a large increase in the experience differential for the U K : a 11% increase in the difference between workers with 0 years of experience and workers with 20 years of experience between 1974 and 1988. In Canada, Kapsalis et al. (1999), among other papers, reveal an 18% increase in the differential between 25 to 30 year olds and 45 to 54 year olds from 1979 to 1989. These increased experience differentials are often interpreted as reflecting the same skill biased demand shifts hypothesized to be behind increased wage-educaton differentials. In particular, they are usually interpreted as revealing an increase in the return to experience over time. MaCurdy and Mroz (1995) examine the evolution of the earnings-age relation-ship from 1976 to 1993 for the US in a cohort framework similar to equation (4.3). Specifically, the authors use CPS data organized into synthetic cohorts to follow the evolution of earnings with age for cohorts defined by individual year of birth. Using CPS data on males for the years 1976 to 1993, MaCurdy and Mroz (1995) show that specification (4.3) is preferred over specification (4.2) but that, rather 70 than increasing slopes for cohort specific wage-age profiles across cohorts, as an increasing returns to experience story would imply, the US data is actually char-acterized by successively lower labour market entry wages and declining slopes of the wage-age profiles for successive birth cohorts. Thus, they find no evidence of rising returns to experience. Moreover, they test and cannot reject a specification in which, for a given educational group, common year effects alter the wages of older cohorts of workers and the entry wages of new cohorts alike. They term this common effect a macro effect and, given their specification, it is the evolution of this macro effect that requires a theoretical explanation. This is quite a different focus from the aim of many of the earlier papers in this literature: to explain ris-ing returns to experience over time. According to MaCurdy and Mroz (1995), one observes rising cross-sectional experience differences because of substantial differ-ences in cohort effects between the pre and post-1976 cohorts. If one were to focus solely on the post-1976 cohorts, the common macro effect specification they find implies a constant slope for the cross-sectional wage-age profile across years. As in earlier, cross-sectional work, MaCurdy and Mroz (1995) find that the patterns they identify are much stronger for less educated males. Thus, the implied macro effects they derive from their specification show almost no trend for males with 16 or more years of education but strong downward trends for all groups with less than 16 years of education. Beaudry and Green (2000) perform a similar exercise using Canadian data covering the period 1971 to 1993, breaking their analysis down by gender and education groups. Again, the largest movements are for high school educated males. For this group, the data period can be effectively split in approximately 1978. Cohorts turning 25 before that year face a pattern in which successive cohorts face higher real labour market entry wages but flatter earnings-age profiles. Cohorts turning 25 after 1978 experience a pattern in which successive cohorts face rapidly falling real entry wages but unchanging earnings-age profiles. The decline in the real entry wages actually accelerates with time and from 1978 to 1993, the weekly earnings of a 25 year old, high school educated male falls by approximately 20% in real terms. Thus, while the pre-1978 cohort experience is best described by (4.3), the post-1978 experience is best described by equation (4.2) with the g function having negative first and second derivatives. It is the acceleration which is behind findings that experience differentials measured in the cross-section have increased dramatically for less educated males in Canada. University educated 71 males face similar sized real declines from the early 1980s to the 1990s and can also be well represented by a constant earnings-age profile with declining intercepts across cohorts. Since the cohort-specific intercepts decline at a constant rate rather than an accelerating rate across cohorts, one finds little evidence of increasing experience differentials using cross-sectional data over time for university educated men. High school educated females show virtually no cohort effects or changes in the slope of the earnings-age profile from the early 1970s to the early 1990s, and university educated females also experience no discernable cohort effects but slight increases in the slope of the earnings-age profile over time. Thus, with the exception of some small movements for university educated females, Beaudry and Green (2000)'s findings for Canada echo those for the US: there is no evidence that any cohort actually obtained increases in the return to experience over the last 20 to 30 years. In contrast to the US results, the post-1980 results for Canada do not reveal significant age-cohort interactions. That is, they do not reveal a common macro effect. Thus, attention for explaining these patterns needs to focus on cohort effects rather than on a macro trend affecting all cohorts equally, as MaCurdy and Mroz (1995) argue is relevant for the US. Gosling et al. (2000) carry out a very similar exercise using U K data on male wages for the period from 1978 to the early 1990s. As in the US and Canadian papers, they find that increasing cross-sectional experience differentials can be described as arising from changes in the rates of growth of cohort effects. As in Beaudry and Green (2000), Gosling et al. (2000) test for age-cohort interactions and cannot reject the restriction that they be excluded from the specification. Thus, in both Canada and the U K after 1978, one can effectively represent wage-age patterns using cohort effects and common across-cohort wage-age profiles (i.e., using specification 1)). Again, this implies the lack of a macro effect as defined in MaCurdy and Mroz (1995) and focusses attention on differences in wage levels across cohorts. In contrast to both the US and Canada, Gosling et al. (2000)'s evidence for the U K shows that successive cohorts have been experiencing real wage increases2. Gosling et al. (2000) also perform a brief exercise in which they relate their estimated cohort effects to a trend, unemployment rates from the time at which each cohort would have entered the labour force and the pupil/teacher ratio in the years before a cohort would have entered the labour force. These covariates 2Findings of increased experience differentials in cross-section data over time arise in this context because of a deceleration in the growth of the cohort effects. 72 explain the movement in the cohort effects very well, with the unemployment rate and pupil/teacher ratios entering with the expected sign. However, simple plots of the cohort effects indicate that the linear trend is probably doing much of the work in explaining the pattern. One potential explanation for the movements across cohorts is that, while the data to this point has been organized by birth cohort, we are actually witnessing an effect related to the type of considerations described in implicit contracting models. Thus, Beaudry and Dinardo (1991) examine US wage data and find evidence that the lowest unemployment rate experienced during a worker's tenure with their firm has significant effects on the current wage level. This approach effectively organizes the data into cohorts defined by the year of entry into a job rather than year of birth and then relates job-entry cohort effects to cohort specific covariates (the lowest unemployment rate during a cohort's time with a firm). In terms of the framework set out above, their specification corresponds to (4.4) without the g and h functions but including a function in both c and t. The latter function is then parameterized as being related to unemployment rates during the tenure of the job. It is noteworthy that Beaudry and Dinardo (1991) find evidence favouring a model in which the important factor for wage determination for a job-entry cohort is the tightness of the tightest labour market during their tenure with a firm rather than just the tightness of the labour market at the start of the job. They argue that this is evidence in favour of implicit contracting models with recontracting. Matching this finding to the results in the birth cohort literature is difficult since Beaudry and Dinardo's environment is one with uncertainty and with wages being on average higher in booms than recessions. In contrast, the birth cohort papers study long term trends that might be predictable by the agents involved. Further, in the case of Canada and the US those trends are negative, leaving no scope for workers to recontract for higher wages in tighter parts of the cycle.3 One goal of this paper is to examine data in which we can observe both job tenure and age to investigate the extent to which the cohort patterns described earlier arise from job specific attributes, as might be suggested in an implicit contract model, versus birth cohort specific factors such as those discussed in Gosling et al. (2000). To the extent that the data fits with a job entry cohort 3Note that the Beaudry and Dinardo evidence is based on the period from the mid 1970s'to the mid-1980s, which corresponds to the period when wages peak and then start down for the less skilled. Since the overall trend in this period is relatively flat, there is still scope to find effects from tightening markets in that data. 73 specification, we would then like to relate movements in average real wages for different entry cohorts to observable, job related factors. A useful paper in this regard is Morissette (1997b), which does not organize the data by job tenure but does relate birth cohort data to job related factors such as the unionization rate, industrial composition and movements in the minimum wage. For the period he studies (1976 to 1996), Morissette finds there have been sectoral shifts (more jobs in the services sector) and a sharp decline in the fraction of young workers who are represented by a union. However, he finds that little of the decline in youth wages can be accounted for by either changes in industrial composition of employment or declines in unionization. He comes to this conclusion based on estimates from a set of standard wage regressions. A key difficulty with this approach is that it attempts to understand movements in the wage distribution by looking only at conditional means. In contrast, we use an estimator that provides conditional (on a set of covariates) estimates of entire distributions. We also extend the analysis into the later part of the 1990s, which Picot and Heisz (1999) argue contains much different labour market patterns compared to the 1980s. Focussing on job characteristics leads to an immediate connection to a series of papers examining the impact of labour market institutions such as minimum wages and unionization on the wage distribution. DiNardo et al. (1996) examine US wage data over the period 1979 to 1988, providing striking evidence of the impact of declines in the real minimum wage particularly for females. They also show that declining unionization plays a substantial role in patterns of rising wage inequality, accounting for 40% of the increases for men. Dinardo and Lemieux (1997) attempt to determine the extent to which increasing dispersion in Canadian and American male wages between 1981 and 1988 can be accounted for by changes in unionization and the minimum wage. Using their measures of dispersion, inequality increases more rapidly in the United States than in Canada. The authors find that 2/3 of the difference in growth can be accounted for by differences in institutions. Over this period, the U.S. experienced decreased unionization, increased dispersion within the union sector and the erosion of the minimum wage. In contrast, in Canada the unionization rate remained fairly constant, the equalizing effect of unions increased dramatically and the minimum wage increased. Institutional features do not seem to account for any of the increase in the wage dispersion that occurred within Canada. Lemieux and Fortin (2000) apply the same type of analysis, though in a less rigorous form, to movements in the Canadian wage distribution from 1988 74 to 1995. They focus, in particular, on the redistributive role of the minimum wage. Donald et al. (2000) develop an estimator for wage distributions conditional on covariates and apply it to examining differences in the wage distributions be-tween Canada and the US in 1989. Like Dinardo and Lemieux(1997) they find that unionization plays an important role in explaining differences, finding a par-ticularly large role for the spillover effects of unions on nonunion wages. These results suggest that it would be interesting to relate movements in cohort specific real wages to cohort specific unionization rates. For Canada, movements in the minimum wage are unlikely to be the primary driving force behind the cohort ef-fect movements because the latter decline continually since the late 1970s while the real minimum wage first declines then increases in Canada over this period. Nonetheless, evidence from these earlier papers suggests that the minimum wage has a potentially substantial impact that needs to be taken into account in the search for the impact of other factors. 4.3 Data In our main empirical exercise, we use data from a collection of Canadian cross-sectional and longitudinal datasets that span the period 1981-98. The main advantage of this data from our perspective is that these are large, representative surveys that include information on job tenure measured as a continuous variable. More specifically, we use the 1981 Survey of Work History (SWH), the 1986-87 Labour Market Activity Survey (LMAS), thel988-90 LMAS, and the 1997 and 1998 Labour Force Survey (LFS). 4 A l l surveys are either from the LFS or supplements to the LFS and are comparable in design and construction5. Each survey is meant 4The year of each survey corresponds to the year for which the questions actually apply. We do not make use of either the 1984 Survey of Union Membership or the 1995 Survey of Work Arrangements because they do not contain a continuous job tenure variable. We do not use the Survey of Labour and Income Dynamics because of difficulties with the comparability of wages recorded in the SLID with those in other datasets. 5One exception to this is that the SWH records whether individuals were union members but not whether they were covered by a union contract if they were not members. All the subsequent surveys recorded union coverage and membership. We believe that the correct manner in which to capture union effects is to measure their impacts on all workers covered by their contracts, whether or not they are union members. To maintain consistency, we considered dropping the SWH from our analysis, but 1981 is a crucial year for allowing us to examine cross-cohort trends given the age grouping used in these datasets. It also allows us to use a set of cyclically similar points at the beginning and end of the 1980s and the end of the 1990s. Instead, we use later year data to estimate a probit on non-union members who are covered by union contracts. We use the estimated coefficients from this exercise, individual covariate values and draws from a 75 to be roughly representative of the Canadian population. In all of the analysis that follows, we use the sample weights included to reflect non-random sampling used in the construction of the surveys. Although the LMAS follows individuals for several years, we treat these data sets as cross-sections. For each survey year, we retain those individuals aged 16-64 who are paid workers, were members of the labour force during November of the year of the survey, and had a high school or less education. We choose to focus on less educated individuals because, as discussed in the previous section, this is the group for whom age related wage movements are the largest in Canada, the U K and the US and thus is the group standing in greatest need of examination. Indeed, in both Canada and the US, there is little evidence of any increase in cross-sectional experience differentials for workers with a post-secondary education. We examine the usual hourly wage from the main job held in November, with the main job defined as the one with the most hours per week. We do not include overtime wages in our analysis. A l l wages are converted to 1998 dollars using the CPI. In the 1997 and 1998 LFS data, years of job tenure is continuous up to 20 but all tenure lengths above 20 are grouped in one category. In much of the work that follows, where we place individuals in job entry cohorts, this grouping at the top is problematic and, as a result, we use only data on individuals with fewer than 20 years of job tenure in each of our sample years. This in turn means that individuals entering the labour market before 1960 and having a long tenure job cannot be represented in any year of our sample, generating a selection bias in estimated cohort effects for those workers. For that reason, we cut individuals who entered the labour market (i.e., turned 25) before 1960. Finally, in January of 1990, the LFS changed the way in which educational information was collected. Prior to 1990, individuals who did not finish high school but completed a post-secondary diploma were grouped with those individuals with some high school education. After 1990 these individuals are included with those individuals that completed both high school and a post-secondary diploma or certificate. As a result of these changes, our educational groups are not perfectly comparable before and after 1990. We include year interactions in our analysis, which, to some degree should control for this compositional change. However, the fact of this seam in the data standard normal distribution to form fitted values of an underlying index function corresponding to union coverage. Those with index values greater than 0 are assigned as covered by a union contract. This selects particular individuals while preserving predicted probabilities of coverage for sub-groups defined by common covariate values. 76 should be kept in mind. 4.4 Basic Patterns in the Canadian Data 4.4.1 Mean Wages As a first step in our analysis, in Table 4.1 we present mean wage patterns by birth cohort and gender in order to establish that patterns documented in earlier papers with other datasets are present in our data.6 To put our discussion in terms of cohort patterns, we use the following grouping: the set of individuals who are aged 55 to 64 in 1981 are called cohort 1; those aged 45 to 54 in 1981 are called cohort 2; those aged 35 to 44 in 1981 are called cohort 3; those aged 25 to 34 in 1981 are called cohort 4; those aged 25 to 34 in 1989 are called cohort 5; and those aged 25 to 34 in 1998 are called cohort 6. We examine average wages for these cohorts in the cyclically similar years of 1981, 1989 and 1998.7 This cohort grouping is obviously not perfect. In particular, cohort 5 will include two years worth of individuals who are also in cohort 3 and cohort 6 includes 1 year worth of individuals who are also in cohort 5. This inexactness is forced upon us by the fact that the data records age grouped in 10 year groupings above age 25 combined with our desire to use cyclically similar years. The inexactness is not large enough to obscure the main patterns we are investigating, however. We will call cohort 4 the new labour market entrant cohort for the early 1980s, cohort 5 the new entrants for the late 1980s/early 1990s, and cohort 6 the new entrants for the late 1990s. Treating 25 to 34 year olds as new entrants focusses on an age after most educational investments have been completed and when individuals are making transitions to more stable working and living arrangements. Comparing outcomes for 25 to 34 year olds in 1981, 1989 and 1998 corresponds to examining patterns at labour market entry across successive cohorts of labour market entrants. In this light, the results in Table 4.1 support conclusions that successive cohorts of labour market entrants are faring worse in terms of wage outcomes. For high school educated men, the 17% decline in real wages for cohort 6 at labour market entry (1998) relative to the average real wage for cohort 4 at 6In calculating the average log wages in Table 4.1 we keep individuals with 20 or more years of job tenure in order to generate results that are comparable to those in earlier papers. The fact that we cut these workers in subsequent estimation does not introduce biases because we condition all our results on narrow job entry cohorts and years of job tenure. 7The unemployment rates in these years were 7.6%, 7.5% and 8.3%, respectively. 77 labour market entry (1981) is of a similar order of magnitude to that recorded in weekly wages across cohorts of high school educated men over the same period by Beaudry and Green (2000). High school educated females also exhibit a decline across labour market entry cohorts (8% between cohorts 4 and 6). The fact that high school educated men experience substantially larger cross-cohort declines in the 1990s than the 1980s fits with the Beaudry and Green (2000) result that cross-cohort declines are following an accelerating pattern for high school educated men. The results support a model with declining entry wages but similar wage-age profiles across successive labour market entry cohorts. Thus, for high school ed-ucated men, cohort 4 experiences a 6% rise in earnings from 1981 to 1989 (i.e., from ages 25-34 to 35-44). Over the same span in the life cycle, cohort 5 (which experiences the shift from ages 25-34 to 35-44 between 1989 and 1998) has a 7% real increase in wages. Similarly, cohort 3 experiences only a 2% increase in real wages from 1981 to 1989 while cohort 4 experiences a 0% increase in wages for a similar age span, running for them from 1989 to 1998. This pattern of similar shaped wage-age profiles shifting down across cohorts is exactly what is found in Beaudry and Green (2000). To reiterate our earlier discussion, the parallel wage-age profiles do not fit with a model with increasing returns to experience. This is true in spite of the fact that changes in relative declines across cohorts lead to the kind of increases in cross-sectional wage-experience differentials noted in the earlier literature. Thus, the wage differential between 35-44 year old high school educated men and 25-34 year old men with the same education rises from 7% in 1981 to 14% in 1989 and then 16% in 1998. As discussed in the previous section, our data includes a continuous job tenure variable. In figures 4.1 and 4.2, we take advantage of that variable to examine job entry cohorts underlying the labour market entry effects seen in Table 4.1. In particular, we plot average real log wages by job tenure, age group and year. Thus, the bottom line in figure 4.1, which focusses on high school or less educated males, shows the average log earnings of individuals who are age 20-24 and have less than 2 years of job tenure in each of our sample years. The other two lines without symbols on them, show the average log wages for older age groups with the same job tenure. The three profiles are strikingly parallel. Thus, this data can be characterized as containing a constant (across years) age differential and falling wages at job entry that affect all age groups in the same way. This suggests that the birth cohort declines displayed in Table 4.1 are actually reflections of 78 declines across job entry cohorts (defined as groups of individuals who start their jobs in the same year), regardless of age. The job cohort declines show up as birth cohort effects in Table 4.1 and in earlier papers because young people are much more likely to be starting a new job at any point in time. The two lines with symbols show average log wages by age group for individuals with 9-10 years of job tenure in the given year. The fact that these follow a roughly parallel pattern relative to the short job tenure profiles for the same age groups suggests a constant (across years) wage-tenure profile. Thus, Figure 4.1 suggests a simple formulation for characterizing the wage data for men with a high school or less education in which individuals face constant age and tenure differentials over time but successive generations of job entrants face substantially lower wages. The constancy of the age and tenure profiles implies that differentials in entry wages between job entry cohorts will persist over time. At first glance, this type of pattern fits with an implicit contract or a dual labour market type of model. Figure 4.2 repeats the plots of figure 4.1 but for females with a high school or less education. The plots for individuals with less than 2 years of tenure again show a roughly parallel pattern, especially after 1987. For females of this education level there is no age differential in wages above age 25. The plots for females with 9-10 years of job tenure match the profiles for individuals with less job tenure in the same age group only roughly at best. Thus, the wage patterns for high school educated women still indicate declining outcomes across job entry cohorts but fit less well with a model with constant wage-tenure profiles than is the case for men. Figures 4.1 and 4.2 only plot wage patterns for individuals under age 45. As discussed in section 4.2, we cut individuals who entered the labour market before the 1960s from our sample for reasons relating to top-coding in the tenure variable. The implication is that individuals over age 45 in 1981 are not in our sample. When we plot similar lines to those in figure 4.1 for males age 45-54, we find that they are again parallel to those for younger age groups after 1986 but have a different pattern from 1981 to 1986. This indicates that cohorts who entered the labour market before 1960 are indeed different from those that follow but that the later cohorts maintain the types of patterns displayed in figure 4.1 even after they age past age 45. For women, there is no such distinction: the lines for individuals aged 45-54 look very similar to those aged 35-44 across the whole range of years. It is instructive to consider the contents of these basic wage plots in the context of the framework set out earlier. Given the findings in the earlier literature for 79 Canada, which are echoed in Table 4.1, we will work with an adjusted version of specification (4.4) that does not include h: w = /* (a) + 9*(P) + A(c) + 7(*) + S(c, t) (4.5) The parallel nature of the lines corresponding to different years of tenure indicate that this specification can be simplified by eliminating the 7 function. Consider, after making this adjustment, the expressions for the mean log wage in year, y, for individuals who are aged a\ and (ai + 1) and who have zero years of job tenure. We will denote the younger person as being from birth cohort bl and the older person as being from cohort 60 • Their respective mean log wages are then given by, wauy = f*(a1)+g*(b1) + \(c1) (4.6) w a i + i , y = / * (ai + 1) + 9*(.b0) + A(ci) (4.7) Note that both individuals are in the same job entry cohort since both start their jobs in year, y. Now consider the mean log wages for individuals of the same two ages in year, y + 1: wauy+l = r(a1)+g*(b2) + X(c2) (4.8) wai+i,v+i = + 1) + <7*(&I) + A(C 2 ) (4.9) If we think of y = 1981, y + 1 = 1989, al = 25 to 35 year olds, and (al+1) = 35 to 45 year olds then these expressions correspond points on the "Age 25-34, 0-2 Yrs Tenure" and "Age 35-44, 0-2 Yrs Tenure" lines at 1981 and 1989. If we take the difference, (4.7) - (4.6) then we get an expression for the wage differential between age groups for new job entrants in year y, (wai+ltV-wauy) = (/*(a1 + l ) - / * ( a 1 ) ) (4.10) + (^(6o)-S*(6i)) + (A(c 1 ) -A(c b ) ) The same differential in year y + 1 is given by, 80 ( W a i + l . y + l — Wauy+l) (r(ai + i)-nai)) (4.11) In figure 4.1, these two differentials (again assigning values to the year and age groups as above) are essentially equal - the actual differentials are .11 in 1981 and .12 in 1989. In a specification such as the one considered here, where there are no job cohort-tenure interactions, this constancy implies one of three possible scenarios: 1) the age profile is constant across time and there are no birth cohort effects once one conditions on job tenure; 2) the age profile is constant across time and there is a constant difference across birth cohorts, i.e., that there has been a linear change in wage outcomes across birth cohorts; 3) the age profile is not actually constant over time but the movements in the age profile slope are exactly offset by movements in the differences between cohorts. The latter option seems quite unlikely because we observe relatively constant age differences between more than one pair of age groups. Thus, the same difference between two given birth cohorts would be required to balance changes in age differentials between young age groups at the start of our data period and changes in age differentials between older age groups at the end of our data period. This possibility cannot be rejected with this data but it seems less plausible than a simple story with a constant age profile. Figure 4.1 also reveals a parallel movement in the mean log wage lines for males with 9 to 10 years of job tenure but different ages. This again implies a specification in which birth cohort effects are either zero or evolving in a linear fashion. However, the constant age differentials witnessed at 9 to 10 years of job tenure are smaller than similar age differentials at 0 to 2 years of tenure. This points to a specification in which returns to age differ by job tenure, with age being much more important for new job entrants. Thus, investigation of the raw data suggests a parsimonious specification given by, In order to investigate the appropriateness of a specification in which wage-age and wage-tenure profiles are constant across time but wages vary across job entry cohorts, we estimate a series of log wage regressions for males and females separately. As in table 4.1, we carry out the estimation using only data from 1981, w = f*(a) + X(c) + 7(t) + X(a,t) (4.12) 81 1989 and 1998 in an attempt to control for cyclical effects. Focussing on these three years is also useful because it allows us to account for something roughly equivalent to birth cohort effects. The grouping of the age variables in these data sets means this cannot be done even roughly for intermediate years such as 1986. In the first specification, we regress individual log wages on a complete set of dummy variables corresponding to age, a set of dummy variables corresponding to two year groupings of job tenure, a full set of year dummies, and the complete set of interactions of all of these sets of variables.8 This provides the most complete representation of age, tenure and year variation possible given the groupings in the data. In essence, this can be thought of as a reduced form version of equation (4.4) where we allow age and tenure effects to vary over time in a manner that would fit with any combination of cohort, wage-age and wage-tenure profiles, including ones that allow the latter profiles to vary over time. After estimating the general reduced form, we estimate two variations of spec-ification (4.12). The first parameterizes the age profile function, f(.), using a set of 5 age dummy variables as defined in footnote 8; the job cohort function, A(.), using a set of 17 dummy variables; the tenure profile function, 7(.), as quadratic in years of tenure; and the tenure-age interaction function, x(.), as an interaction of years of tenure with the dummy variables for age. To create job entry cohorts, we first construct the job start year using the sample year of the observation and the recorded job tenure. We then group start years into two year cohorts begin-ning in 1960. Thus, job entry cohort 1 consists of individuals whose jobs began in 1960 or 1961, job entry cohort 2 consists of individuals whose jobs began in 1962 or 1963, etc.. We use two year groupings to allow for detail in cross-cohort movements while imposing extra smoothing relative to the variability we would get with single year entry cohorts.9 In our first variation on specification (4.12), we 8The age variables are grouped into dummy variables corresponding to ages: 16-19, 20-24, 25-34, 35-44, 45-54, and 55-64. The job tenure dummy variables correspond to the following year groupings: 0-1, 2-3, 4-5, 6-7, 8-9, 10-11, 12-13, 14-15, 16-17, 18-19. To maintain comparability with the 1999 LFS, where job tenure is top coded at 20 years, we drop observations with 20 or more years of job tenure in every sample year. The calendar year dummy variables correspond to 1981, 1989 and 1998. 9The one exception to this classification rule is that we form one three year cohort grouping for the years 1992, 1993 and 1994. We do this in order to create a last tenure cohort consisting of job entrants in the years 1997 and 1998. Without this adjustment, our last cohort would consist only of 1998 entrants. Since all other job cohorts are observed for at least two years, a single year grouping for the last cohort makes it non-comparable with all the earlier cohorts. The 1999 version of the LFS is now available and would help us to rectify this problem. However, the industry codes in the 1999 LFS do not match codes in earlier years, making the 1999 LFS 82 also include interactions of the job cohort dummy variables with years of tenure. This allows for macro effects as defined by MaCurdy and Mroz (1995): common shocks affecting all workers, regardless of tenure or age, in a given year. We allow for this possibility because a cursory glance at figure 1 might suggest that such effects exist; although that impression fades once one concentrates on long term trends by picking out cyclically similar points. The second variation we estimate is specification (4.12) itself, without the cohort-tenure interactions. Both variations on specification (4.12) represent considerable restrictions on the first, general specification. A formal test of the restrictions implicit in moving from the first specification to the restricted specification including cohort-tenure interactions indicates we cannot reject the restrictions at the 1% level of signifi-cance. For females, the same test indicates a rejection of the restrictions at the 1% level, though the test statistic is quite close to the critical value.1 0 The test of the restriction that the interactions of job entry cohort with tenure are jointly insignif-icant cannot be rejected at the 1% level for males but can be rejected for females. Again, though, the test statistic value is close to the critical value for females.11 For females, the adjusted R-squared statistics reveal that the restrictions perform quite well in spite of their formal rejection. In particular, the adjusted R-squared values are .192, .189 and .187 for the general specification, the specification includ-ing cohort-tenure interactions, and specification (4.12), respectively. This suggests that the restricted specification performs nearly as well as the unrestricted speci-fication in describing the variation in the data. As figure 4.2 suggests, the relative fit of the restricted model is not quite as good for females, but it still appears to capture much of the variation depicted in the unrestricted regression. One point of interest is to delineate between a model that includes only job cohort effects and one that includes both job and birth cohort effects. Unfortu-nately, collinearity among the age, birth cohort, tenure and job cohort variables is unusable for our purposes. 1 0The test statistic for males is distributed as F(57, 26359) and equals 1.13. For females the test statistic is distributed as F(55,21956) and equals 2.07. The relevant critical value at the 1% level of significance is approximately 1.48 in each case. n The test statistic for males is distributed as F(17,26416) and equals 0.92. For females the test statistic is distributed as F(17,22011) and equals 5.25. The relevant critical value at the 1% level of significance is approximately 1.97. We also tested the restrictions implicit in moving from the first specification to specification (4.12) without cohort-tenure interactions. For males the relevant test statistic is distributed as F(74,26359) and equals 1.08. For females the test statistic is distributed as F(72,21956) and equals 2.83. The relevant critical value at the 1% level is approximately 1.41. Thus, we can again reject the restrictions for females but not for males. 83 sufficiently strong that it is difficult to see anything significant in a specification including both job and birth cohort indicator variables. We cannot formally reject the restriction that the birth cohort variables do not enter in a specification such as (4.12) but, again, the adjusted R-squared changes very little whether they are in or not. Thus, the most we can say is that a simple specification including constant wage-tenure and wage-age profiles and job cohort indicators fits the data well and little is added by considering birth cohort variables as well. Thus, we focus our attention on factors that would drive differences across job cohorts. Note that, given the observations based on figure 4.1, any birth cohort effects would have to enter in a restrictive manner. Table 4.2 contains estimates of the age and job tenure profiles and a selection of the job entry cohort coefficients from the restricted regressions. The estimates confirm what is observed in figures 1 and 2. The age profile for men shows sub-stantial increases in wages for men from the age 16 -19 group through to the age 35-44 group, with flattening thereafter. The age profile for women is much flat-ter, particularly after age 25. On the other hand, the tenure profile for women is steeper and has more curvature. The job entry cohort dummy variable coefficients indicate that new job entrants in 1989 earned 11% less and new job entrants in 1998 earned 20% less than new entrants in 1981. For females the decline from 1981 to 1989 new job entrants is 7% while the decline from 1981 to 1998 is 13%. Thus, the job entry cohort declines for women are about half those for men and both experience a relatively constant decline across the period. Having established that a restricted specification with constant age and tenure profiles and job entry cohort effects does an adequate job of describing the variation in the data, it is worth examining what those profiles look like. Figure 4.3 shows the average log wages in 1981, 1989 and 1998 for a 1981 job entrant male who is age 25-34 in 1981 and who keeps his job over this period. 1 2 Similarly, it follows 1989 and 1998 new job entrants through time. Note these are simple sample averages and thus are not constructed incorporating any restrictions, such as those imposed in specification (4.3. In the same figure, we plot the fitted profiles for each of these cohorts of job entrants constructed using the coefficients in Table 4.2 and thus reflecting the restrictions that age and tenure profiles are constant across 12Thus, to construct the profile labelled "1981 Job Entry Cohort", we calculate the average log wage of males aged 25 to 34 with less than 2 years of job tenure in 1981. We also calculate the average log wage of individuals aged 35 to 44 with 8 or 9 years job tenure in 1989 and the average log wage of individuals aged 45 to 54 with 17 or 18 years of job tenure. 84 time. The figure gives a good picture of the worsening outcomes for subsequent cohorts of new job entrants. Focussing on the simple sample averages, it is clear that the 1989 new job entrants start at substantially lower wages (8% lower) and experienced similar wage increases over the first 9 to 10 years of their job. The 1998 entrants start another 13% lower. The fitted averages based on specification (4.4) show very similar patterns. A key point from figure 4.3 is that successive cohorts of job entrants of high school or less educated men are faring substantially worse in terms of entry wages and show no evidence of catching up. Figure 4.3 also reveals why job cohort-tenure interactions did not enter our restricted specification significantly. The growth in wages for the first 8 to 9 years on the job is very similar for the 1981 and 1989 entry cohorts. This means successive cohorts of job entrants can expect to experience the same wage growth on the job but starting from lower initial wages. If the lower initial wages reflected a macro shock affecting workers at all tenures, one would see a flatter wage-tenure profile slope for the 1989 than the 1981 cohort. One would also observe significant cohort-tenure interaction terms in our regressions. The fact that neither of these are observed means we need to focus on factors driving differences across cohorts rather than common macro effects. Figure 4.4 presents the same job entry cohort plots for females. As for males, the simple averages show declines across cohorts and quite similar wage-tenure profiles for the first 9 to 10 years on the job. Although the declines are not as large, this figure also paints a depressing picture of worsening outcomes for new job entrants with no evidence of catching up with future time on the job. 1 3 4.4.2 Covariates The pattern of declining wage outcomes across job entry cohorts has parallels in the distribution of other work related characteristics across cohorts. In table 4.3, we present the industrial distribution, union coverage rates and real minimum wages for three job entry cohorts in various years. Thus, for the 1981 job entry cohort, we examine covariate distributions for individuals who are age 25 to 34 and have less than 2 years of job tenure in 1981. We follow them through time on the same job. To do this, we construct a synthetic cohort by examining individuals who are 35 to 44 years old and have 8 to 9 years of job tenure in 1989 and individuals who are 13Beaudry and Green (2000) argue that this type of pattern is not consistent with a skill biased technical change inducing increasing returns to skills since such increased returns should show up as steepening profiles across successive cohorts. 85 45 to 54 years old and have 17 to 18 years of job tenure in 1998. The 1989 cohort is similarly followed through time in a manner meant to mimic the experience of someone who enters a job in 1989 and keeps it through 1998. For males with high school or less education, shown in the top of the table, following the 1981 job entry cohort across cyclically similar years shows an industrial distribution that shifts toward manufacturing, transportation, health and education, and public administration. These changes are offset by relatively dramatic shifts away from primary, construction, trade and service type jobs. Since the entries are intended to follow individuals who stay on a job, the changes in the industrial distribution reflect differing probabilities of job termination with time on a job, not movements into new sectors. In this sense, it is not surprising that jobs that are typically more seasonal in nature, such as construction and primary sector jobs, are much more strongly represented just after job entry than after 9 or more years of job tenure. For our purposes, what is most of interest is to see how the distribution at job entry and movements with increasing job tenure vary across job entry cohorts. In this regard, the key movements are substantial increases in the proportion of workers in service and trade sector jobs offset by declines in primary, construc-tion and transportation jobs. This represents movements toward lower paying jobs across new job entry cohorts. Interestingly, there is virtually no change in the proportion in manufacturing jobs. These industrial shifts are accompanied by large movements in the proportion of new job entrants covered by collective bar-gaining agreements. Thus, new job entrants in 1981 have a coverage rate of 41% and this rises for members of this cohort who remain on the same job. However, by 1998 new job entrants in the same age group have a unionization rate of only 17%. The Coverage Rate in Industry Variable represents the unionization rate in the individual's 3 digit industry. The fact that this does not decline as rapidly as the individual coverage rate for the new entrants indicates that the cross-cohort decline is not just reflecting general declines in unionization: declines are concen-trated among new entrants. This observation along with patterns in the industrial compositions potentially fits with a model in which job holders who entered their jobs at earlier dates maintain access to good union jobs in higher paying sectors while newer job entrants are relegated to non-union, service and trade sector jobs. Finally, the minimum wage variable shows a pattern of decline between the 1981 and 1989 new entrants but recovers over half that decline by the time the 1998 new entrants begin their jobs. Thus, while minimum wages may help explain part 86 of the 1980s cross-cohort decline, they cannot help explain the continuing fall in the 1990s. They do move enough, however, that we would like to control for them in order to get a clearer picture of the effects of other covariates. The patterns for women shown in the bottom of Table 4.3 are somewhat differ-ent from those of men. As with men, the proportion of jobs in the services sector declines with time on the job and this is offset by increases in public administration and transportation. However, in contrast to men, the numbers in the primary and construction sectors are too small to show much of a pattern and women actually move out of the manufacturing and financial services sectors over time. Again, there are substantial increases in the proportions in service and trade sector jobs across entry cohorts. For women, these are offset by declines in access to public administration and financial services jobs. As with men, the union coverage rate falls dramatically across cohorts, from 22% for new job entrants in 1981 to 11% for new entrants in 1998. Overall, women also show a pattern, though somewhat more muted than for men, of more recent job entrants having reduced access to union jobs in higher paying sectors. 4.4.3 Estimation: Mean Regression Models We turn now to combining the insights on wage patterns across cohorts with those on job characteristics across cohorts. Since the evidence in section 4.4 indi-cates that constant age and tenure profiles are a reasonable summary of the data patterns, we will focus on the impact of controlling for covariates on the cohort specific tenure profile intercepts. Declines in these intercepts, which capture de-clines in average wage levels across cohorts at any age and tenure level in our specification, form the main movement of interest in our data. To carry out this investigation, we estimate a series of mean log wage regressions using data from all our sample years. We then examine what happens to the cohort specific intercepts as other covariates are added. Table 4.4 contains estimates of the cohort specific intercepts from a series of specifications. The first specification mimics that in Table 4.2 but using all the years of data instead of just 1981, 1989 and 1998. As in the specification from Table 4.2, this specification includes dummy variables corresponding to age groups as well as years of job tenure, job tenure squared and interactions between age groups and job tenure. To control for differences across the business cycle we include the national unemployment rate as an additional regressor. Coefficients 87 relating to covariates other than the cohort indicators are presented in Appendix table A . l . The patterns in the first column of Table 4.4 are very similar to those we have described in earlier sections. In particular, the 1980/81 entry cohort and 1988/89 entry cohort effects (each measured relative to the entry wage level for the 1997/98 cohort) are highlighted and show a decline of 22% from the 1980/81 to the 1997/98 entry cohort and a decline of 12% from the late 1980s to the late 1990s. This slightly accelerating decline fits with patterns discussed earlier. The other entry cohort coefficients suggest a relatively steady decline within each decade. The second column in Table 4.4 reports entry cohort effects when additional job related regressors are added to the specification. In particular, the second specifi-cation includes dummy variables corresponding to each province, dummy variables relating to two digit industries, a dummy corresponding to whether the individual is covered by a collective bargaining agreement, and a variable corresponding to the unionization rate in the individual's three digit industry. The latter variable is included to capture union spill-over effects, where it is anticipated that industries with higher unionization rates will have higher nonunion wage levels because of union threat effects. As noted in the previous section, the distribution of these covariates changes dramatically with job tenure. One way to think of this spec-ification is to start from the assumption that job specific characteristics do not change during the duration of a job. Thus, we could define job entry cohorts not just by the start date of the job but by immutable characteristics of the job holder (i.e., age, education and gender) and immutable characteristics of the job (i.e., union status, industry and province). This would result in a very large number of job "cohorts", each with its own wage-tenure relationship. Our specification imposes the restriction that these wage-tenure profiles have the same slope and curvature across all of the cohort groups for a given gender but that the height of the profile can vary according to the year the job started (i.e, what we are calling the job entry cohort), age, union status, industry and province. The profile itself may reflect a combination of the average profile we would observe if all job starters kept their jobs and selection effects as a non-random group of job holders experi-ence a job separation each year. Interestingly, the analysis in section 4.4 indicates that this hybrid profile can be reasonably described as constant across age groups and job entry cohorts. After controlling for job specific characteristics, the cross-cohort declines are reduced considerably: by approximately a third from the early 1980s cohorts to the 88 late 1990s cohorts and by a half from the later 1980s to the late 1990s cohorts. This indicates that the changing patterns of job characteristics across cohorts observed in Table 4.4 played a substantial role in the declining outcomes of subsequent cohort of job starters. This is particularly true of the decline in the 1990s. Specification 3 replicates the specification in the second column but allows for differing cohort effects, age profiles, tenure profiles and industry and province effects by union status. The effects of covariates other than job entry cohort, shown in Appendix table A l , indicate that the tenure profile is considerably flatter in the union sector, in accord with earlier studies of union impacts. Interestingly, the overall decline across cohorts is nearly identical in the union and non-union sectors: about 15% from the 1980/81 cohort to the 1998/99 cohort. However, the pattern over the 1980s and 1990s is quite different. The decline occurs almost entirely across 1980s cohorts in the nonunion sector but mainly across the 1990s cohorts in the union sector. 4.5 Estimation of Conditional CDFs 4.5.1 An Estimator for CDFs Conditional on Covariates Our ultimate goal in this paper is to arrive at a better understanding of the source of cross-cohort movements in wages. We are interested in that, in turn, because of its contribution to movements in overall wage inequality. Given that goal, studying movements in conditional means can provide only a partial account of cross-cohort wage differences and their contribution to inequality. In particular, the distributions of wages for successive cohorts of job entrants may differ in their shape as well as their mean, and in a way that either accentuates or reduces the impact of cross-cohort differences on overall inequality. Further, we can understand only part of the impact of covariates such as union status on inequality trends if we restrict our attention to decompositions of conditional mean wages. For these reasons, we turn to implementing a flexible estimator of cumulative distribution functions (CDFs) with our data. This estimator allows us to observe the impact of covariates such as union status on the CDF and then to perform decompositions of the overall movements in the wage distribution. The estimator we use is presented in Donald et al. (2000). Here we provide a brief overview of its main features. The estimator takes advantage of advances in the large literature on estimating hazard functions to implement a very flexible 89 specification with covariates having different effects over different portions of the wage range. In particular, we directly estimate hazard functions using our wage data and then perform simple transformations to obtain estimates of corresponding CDFs. The result is an estimator that introduces covariates in a very flexible and tractable way while still meeting the basic requirements that corresponding wage densities integrate to one and that the conditional probability of observing wages in any given range takes a value between 0 and 1. Hazard functions are typically used in situations where the dependent variable is expressed in units of time, and in that situation the coefficient estimates bear a natural interpretation. In our case, the coefficients do not have as natural an interpretation and we prefer to view the hazard function as simply a convenient, flexible function that allows us to generate estimated CDFs that respect basic rules about probabilities. For that reason, we will not present or discuss our coefficient estimates directly but instead will present various conditional CDF plots. To understand how the estimator is implemented and the assumptions behind it, we will begin with the definition of the hazard function of a random variable, Y , conditional on a set of covariates, X : h(y/x) = ^ \ (4.13) S(y/x) where S(y/x) is the one minus the CDF of Y and f(y/x) is the conditional density function. A common empirical specification for the hazard function is a propor-tional hazard specification in which the conditional hazard is modelled as the product of a baseline hazard function, ho(y), which is common to all observations regardless of their covariate values, and a term which depends on the covariates. Thus, h(y/x) = h0(y)exp(xP) (4.14) where, 6 is a vector of coefficients to be estimated. The flexibility of a proportional hazard estimator in capturing different shapes of the hazard (and thereby the density) function depends in part on the form of the baseline hazard. We use a step function for the baseline hazard with the intervals in the step function (which we call baseline segments) chosen to be $0.20 in width between the 5th and 95th percentiles of the unconditional (on any covariates) hourly wage distribution, and wider in the tails of the distribution1 4. Flexibility 1 4 Donald et al. (2000) provide a formula for calculating appropriate baseline segment widths 90 in covariate impacts is introduced by partitioning the support of y into p = 1,...P sub intervals and allowing for a different 3 vector, which we will term Bp, for each sub-interval. We refer to these sub-intervals, which in practice are constructed to be subsets of the complete set of baseline segments, as covariate segments. We allow for 6 covariate segments defined approximately by the 15th, 30th, 45th, 60th and 75th percentiles of the unconditional wage distribution.1 5 To obtain covariate conditional CDF estimates we first estimate the likelihood function corresponding to the probability of observing our wage and covariate data, with the likelihood function constructed based on the hazard function as is standardly done in duration literature. We recover estimates of the Bp vectors and the parameters defining the baseline hazard. We then generate estimates of the survivor function corresponding to a wage in the j*th baseline segment using the well known relationship between the survivor function and the hazard, pij*) S(yj*/x) = exp{- J2 exp(xPk)Ak) (4.15) fc=i where, p(j*) is the covariate segment corresponding to the j*th baseline segment and Afc is the integral of the baseline hazard across the whole covariate segment for every covariate segment to the left of the one containing j * and from the left end of the covariate segment to j * in the covariate segment containing j * . Once we have the estimate of the survivor function, the CDF estimate is constructed as one minus the survivor function. Since the underlying estimation is maximization of a well defined likelihood function and the CDF is just a function of the estimated parameters from that maximization, we can obtain standard errors for the CDF estimate at any point using the Delta method. Donald, Green and Paarsch(2000) discuss the asymptotic properties of this estimator and the constructed standard errors in some detail. The estimator as described to this point allows for the incorporation of stan-dard covariate effects on the entire distribution. However, we are also interested in capturing the effects of minimum wage changes, which will have direct impacts that is similar to optimal binwidth formulae for histograms. We use that formula but then rounded our interval size to an even 20 cents. 1 5 Donald et al. (2000) present a Monte Carlo investigation of the estimator that suggests that this number of covariate segments does a good job of capturing complex covariate effects. In particular, with similar sample sizes to ours, they find that increasing the number of covariate segments from 5 to 20 adds little to the ability of the estimator to capture complex patterns. However, the estimator performs markedly worse with fewer than 5 covariate segments. 91 on the shape of specific parts of the distribution. Given findings in DiNardo et al. (1996) that minimum wage movements had substantial effects on the US wage distribution over the 1980s, we want to control for potential impacts of the consid-erable movements in the minimum wage in Canada over our time period. Following Green and Paarsch (1996), we allow for direct minimum wage effects by including a dummy variable which takes on a value of one if a given baseline segment contains the minimum wage that is relevant for the observation given the year and province, and zero for all other baseline segments. A minimum wage coefficient is introduced so that the hazard function for individual i in segment j is ^jexp(xi(3p^ + <5) if the minimum wage falls in segment j (where jj is the parameter corresponding to the jth step in the baseline step function). For a segment j ' which does not hold the minimum wage, the hazard function is given by r)jexp(xi[3p^). In effect, this spec-ification allows for a spike in the distribution at the minimum wage and is in the spirit of a time varying covariate in standard duration models16. In addition to allowing for an effect at the minimum wage we allow for the minimum wage to influence the distribution at wages near the minimum wage. This allows for pos-sible spill-over effects above the minimum wage and reduction in the probability of seeing wages below the minimum. In particular, we define dummy variables for the following wage ranges: $0.11 to $0.30 below the minimum wage; $0.31 to $0.70 below the minimum wage; $0.71 and more below the minimum wage; $0.11 to $0.50 above the minimum wage; $0.51 to $1.10 above the minimum wage; $1.11 to $2.10 above the minimum wage; and $2.11 to $3.11 above the minimum wage. The precise manner in which we obtain identification using this scheme is described in Green and Paarsch(1996). Intuitively, we assume that minimum wages have no spillover effects on the wage distribution above $3.10 above the minimum wage. Movements in the minimum wage then cause movements in the actual dollar val-ues that are assumed to be affected by spillover effects.17 The minimum wage is 16Since we work with $0.20 intervals, a more precise description would be to say that we allow for an increase in the probability of earnings a wage plus or minus $0.10 around the minimum wage. 1 7The intuition behind our estimates of the minimum wage is in some way similar to that in Meyer and Wise (1982) and Dickens et. al.(1997). The main difference is that our estimator allows for a flexible, semi-parametric estimator of the underlying wage density and thus does not rely on functional form assumptions related to the density to help in identification. This is possible in our context because we have many values of the real minimum wage to use in our identification. Our estimator also differs from Meyer and Wise(1982) and DiNardo et al. (1996) in that it allows minimum wage changes to affect the wage density above the minimum wage. DiNardo et al. (1996)'s approach is essentially a decomposition method, pasting the part of the density at and below the minimum wage from the wage density from the early part of their period 92 set at the provincial level in Canada and displays considerable variation in our time period, including a period of real decline in the early and mid-1980s in most provinces, a period of near perfect stability for the real minimum wage in the late 1980s in Ontario and Quebec, and a period of real increases in many provinces in the 1990s. The considerable variation in changes in real minimum wages among provinces over time provides substantial power for identifying minimum wage im-pacts on the distribution1 8. The ability to directly control for minimum wage effects is a major advantage of this estimator over competing approaches such as quantile regression. We implement this estimator separately for the male and female high school educated samples we used in the previous sections. We use the same set of covari-ates as are used in the third specification in Table 4.4. Thus, we include a set of age dummy variables, a set of dummy variables corresponding to job entry cohort, tenure, tenure squared, interactions of the age dummies and tenure, province and industry variables, the union coverage rate in the individual's industry, and the unemployment rate. Interactions of all of these variables with a dummy variable corresponding to being covered by a collective agreement are also included, thus allowing for flexible union effects. 4.6 Results 4.6.1 Estimated Marginal Covariate Impacts As discussed earlier, the estimated parameters of this model do not bear easy interpretation. Thus, rather than presenting and discussing the parameters di-rectly, we will present a series of fitted wage density plots constructed from our estimation in order to portray the marginal effects of various important covariates. Density estimates are obtained by differencing the CDF estimates described ear-lier. Our density estimator is a nonlinear function of covariates and parameters and thus, marginal effects will be a function of the covariate vector value as they are, for example, in probits. We will demonstrate marginal effects relative to a base covariate vector value which corresponds to a 25 to 34 year old individual in 1981 (i.e., with 1981 values for the minimum wage in Ontario, facing the 1981 un-onto the wage density from the end of their period, rather than an estimator of the minimum wage effect. 1 8 For the period 1981 to 1990, there are over 50 separate adjustments to provincial minimum wages, with a further 20 occurring from 1991 to 1998. 93 employment rate, and with the 1981 job entry cohort dummy variable turned on) with less than two years job tenure who is in a nonunion job in the manufacturing sector in Ontario with an average value for the unionization rate in their industry. In the first set of figures, this individual is male. In figure 4.5 we plot the fitted density for our base case individual both with the dummy variables related to the minimum wage set to values based on the 1981 Ontario minimum wage, which we will call wmin, and with the minimum wage dummy variables all set to zero throughout the wage range. Thus, we plot the density with and without the minimum wage. We do not view the plot with the minimum wage dummies set to zero as an accurate depiction of what the wage density would look like in the absence of a minimum wage since the variation in the minimum wage we work from does not approach an elimination of the minimum wage. Rather, we use the case with the dummies set to zero as a benchmark to showcase the impacts of the minimum wage on the distribution. The results show that minimum wages reduce the density below the minimum wage, induce a spike right at the minimum wage and also have substantial spill-over effects in the form of added mass in the region up to approximately $3 above the minimum wage. The fact that the remainder of the density is relatively unaffected partly arises from our assumptions since we do not allow the minimum wage to affect the hazard rate above $3.10 above the minimum wage. However, we tested down from a specification allowing the minimum wage to have effects farther up the wage range and found that we could not reject the restriction that it had no effect beyond $3 above the minimum. The effects on the density up to $2 above the minimum wage are all statistically significantly different from zero at the 1% significance level. Using this estimated pattern, a reduction in the minimum wage would lead to a reduction in mass from the old minimum wage up to approximately $3 above the old minimum, an increase in mass at the new minimum wage and in the wage range between the old and new minima, and no change either below the new minimum or above $3 beyond the old minimum. Thus, a reduction, such as those that occurred through the 1980s in many provinces, would have the effect of flattening the lower portion of the wage distribution. However, there is only a small part of the total mass of the male wage distribution in this region. For this reason, minimum wage changes are unlikely to be a substantial driving force behind the movements we are studying. In figure 4.6 we again plot the density for the base case individual, but also 94 plot the density for an individual with all the same characteristics except that his job is covered by a collective agreement. The union wage density is predictably shifted to the right relative to the nonunion distribution. The union density is also more symmetric, with the two densities having similar right tails but the union wage density having a thinner left tail, and the minimum wage having more impact on nonunion wages. This fits with claims that unions disproportionately support the wages of low wage workers. However, the differences in skewness between the two densities is not as exaggerated as is often reported for US wage distributions. Donald et al. (2000) argue that this occurs because higher union concentrations lead to larger union threat effects and thus more union-like wage distributions in the nonunion sector. To examine this point, in figure 4.7 we plot the densities for our nonunion base case person with the union concentration in their industry set at the 25th percentile value for the sample (0.40), the median value (0.43) and at the75th percentile value (0.59). The differences between the lowest and highest coverage rate densities are weakly reminiscent of the differences between the union and nonunion densities in figure 4.6, where union concentration in the industry was held constant. Thus, higher union concentration in the industry leads to a nonunion wage distribution that is both shifted to the right and more symmetric. Overall, direct union membership effects are substantial but not evenly distributed across the wage range. Indirect union effects tend to be much less pronounced. In figure 4.8, we replot the density for our base case and then for an individual identical to the base case except that he is employed in the service sector rather than the manufacturing sector. The differences are substantial, with the service sector wage density having a much higher minimum wage spike and much lower mass in the $13 to $23 range. As with the union effect estimates, the industry effects are substantial and differ across the wage range. Finally, in figure 4.9, we replot the base case density and then plot densities for individuals identical to the base case individual except that they have the 1989 and 1998 job entry cohort dummies turned on respectively. Since all other covariates are held constant, including values for the minimum wage variables and unemployment rate, the differences across these densities represent the pure cohort effect estimates. Above approximately $18, the 1989 and 1998 entry cohort densities appear as leftward shifts of the 1981 density. Below that point, the 1989, and especially the 1998, densities have much greater mass and greater minimum wage spikes. Thus, the 1998 density appears as the 1981 density would if it were 95 pushed to the left but got backed up against the wall imposed by the minimum wage in the process. To assess this, we replotted figure 4.9 but with the minimum wage variables all set to zero throughout the wage range in figure 4.10. While there is still extra mass at and below $10 in the later period densities, the shift relative to the 1981 density does appear more symmetric. Thus, while changes in the minimum wage do not appear to have large enough effects to account directly for much of the movement in the wage distribution, the existence of a minimum wage has an impact on how changes generated from other sources are distributed. Figure 4.11 replots the figure showing pure cohort effects (figure 4.9) for women. The figure clearly reveals a much larger minimum wage spike for females than males, which arises because the female densities are shifted substantially to the left relative to their those of their male counterparts. Minimum wage impacts on the shape of the density are similar to those depicted for males in figure 4.5 but are larger in scale and persist farther up the wage range. Impacts of unionization, changing the industry union coverage rate, and changing industries are also very similar in nature to those depicted for males. For brevity, we do not present the female counterparts to figures 4.5 to 4.10. Returning to figure 4.11, one again observes a shift left in the density for successive job entry cohorts. However, in contrast to the outcome for males, the upper tail of the distribution (above $15) does not change substantially. Rather, the decline in average wages across cohorts for females that was presented earlier arises because of reduced mass in the mid-wage range and substantially increased mass at and near the minimum wage. 4.6.2 Decompositions of Shifts in the C D F We next pursue a decomposition exercise in order to summarize the impacts of the covariate shifts seen in Table 4.1 on the cross-cohort movements in the wage density. To construct the decompositions, let us first write the estimated CDF for entry cohort, c, in a given year, t, as a nonlinear function of the estimated parameters, /?, and the covariate matrix relevant for cohort c in year t, Xct . Thus, the CDF equals F(/3,Xct). To construct this estimated CDF we select the subsample of individuals observed in cohort c in year t. We then construct an estimated CDF for each individual in the subsample using the parameter (3 and the formulae related to our estimator described earlier. Finally, we average across this set of estimated CDFs at the midpoint of every baseline segment. It is useful for expositional purposes for us to partition the Xct matrix into columns relating 96 to the cohort dummies, Dc, the dummies related to the minimum wage in year t, MWt, the collective bargaining dummy variable, Ua, the union concentration rate in the individual's industry, CONCct, and all other covariates, Xoa- Thus, we can rewrite the CDF function as F(B, Dc, MWU Uct, CONCct, X0ct)-We first plot the percentage difference between the CDF for the 1981 entry cohort in 1981 (i.e., c=81 and t = 81) and the CDF for the 1998 cohort in 1998 (i.e., c=98, t=98). In the terminology set out above, this equals F(B, D98, MW98, U98$8, CONC98>98, X 0 9 8 > 9 8 ) - F'B, Dsi, MW81, U81,S1, CONC8h8u X 0 8 i , 8 i ) - In figure 4.12, we plot, for males, this fitted difference along with the difference estimated using CDFs for the 1981 entry cohort observed in 1981 subsample and the 1998 entry cohort observed in 1998 subsample constructed from simple histograms in which the bins are constructed to be the same as the baseline segments in our estimator. The two lines in the figure are extremely similar, indicating that our estimator performed very well in fitting the relevant CDFs and their differences, even though it was not fitted directly to these differences. The differences in the CDFs for wages for new entrants in 1998 and 1981 have a flattened U shape. In particular, proportional real wage declines from the 1981 to the 1998 cohorts are substantially smaller below the 10th percentile than for the rest of the distribution. From approximately the 25th percentile to the 75th percentile the proportionate wage losses are large (slightly over 30%) and relatively equal, and then above the 80th percentile the losses are again somewhat smaller, though not nearly as small as in the left tail. This implies that comparing across job entry cohorts at the same education level, the within group wage dispersion actually declined between 1981 and 1998. In figure 4.13 we decompose the contribution of observable variables to the overall differences in CDFs between cohorts displayed in figure 4.12 for males. The fitted difference between the two distributions is shown by the line labelled Diff0. As a first step, we recreate the CDFs for each member of the 1998 cohort sample using their own covariate values but setting the minimum wage dummies to values appropriate for the 1981 minimum wage in the individual's province. Averaging across these individual CDFs yields the fitted CDF, F(B, D98, MW8i, 9^8,98) CONCcgs$8i -Xo98 ,9s)- The line marked with squares in figure 4.13 is the constructed difference, Diffi = F(8, D98, MW81, U98i98, CONCc98t98, X098,98) -F(B, D81, MW8i, <78i,8i, CONCc81>81, X08i:si). Comparing this line to the basic fitted distribution line, Diffi shows the contribution of the change in the real min-97 imum wage from 1981 to 1998 to the differences in the wage distributions between 1981 and 1998 new entrants. These lines differ only slightly at and below the 15th percentile and not at all above the 20th percentile, confirming our projection that real minimum wage changes play only a minor role in the cross-cohort decline in wages.19 We next attempt to assess the importance of changes in unionization across cohorts. To do this, we first fit a simple probit on union status using 1981 data on the 1981 entry cohort. We regress union status on all the covariates used in our wage density estimates except for the union status variable itself and the indus-try union concentration variable. We take the parameters from this estimation to represent the unionization process for new entrants in 1981. We use these coeffi-cients to construct fitted estimates of the probability that a member of the 1998 new entrant sample would be unionized if they had faced the 1981 unionization process. We replace the individual's actual union status with this fitted status variable and then construct individual CDF estimates, again using the 1981 min-imum wage. Averaging across these individual CDFs yields the estimated CDF, F(J3, A )8 , MW8i, Ug8fiu CONC^ss, ^ O Q S . O S ) . The difference between this CDF and the estimated 1981 CDF is called Dif f2 . Comparing Diff2 to Diffi shows the marginal contribution of the change in the unionization process over time to the cross cohort wage decline. Note that since the unionization process was fitted using the same covariates used in the density estimation, this marginal impact is the pure impact of the change in the unionization process itself. This would be different from the total impact of changes in unionization on the wage distribution since the total change would include changes that occur because 1998 entrants have different covariate values that would imply different unionization rates even if the unionization process has not changed. Defining the marginal union effect in this way allows us to separate it from effects of other changes such as the change in the industrial composition. Interestingly, figure 4.13 shows that the declining likelihood of unionization over time has only marginal impact below the 25th per-centile of the wage distribution but more substantial effects in the upper half of the distribution. This fits with the idea that the lowest paying jobs are not union jobs in any year and thus the decline in unionization hits those farther up the distribution harder. It is worth keeping in mind that we are looking at a relatively low educated group when we make this statement. 1 9 As noted earlier, the minimum wage had largely recovered in real terms to its previous level. 98 We next construct a set of CDFs in which we use the 1981 entry cohort sample but set their cohort dummy variables to the 1998 entry cohort value. Averaging these yields, F(0, Dg8, MWSi, C/ 8 1 > 8 1 , CONCc8i,8i, X08i,8i) and the difference be-tween this and the original CDF difference yields Diffo. A comparison of Diffa and Dif f2 shows the marginal impact of changes in the remaining covariates. Note that this impact occurs both through their direct impacts on the wage density and through an indirect impact operating through the unionization process. The re-maining covariates are provincial dummy variables, the unemployment rate and industry dummy variable. The unemployment rates for 1981 and 1998 are nearly identical, so the observed effects are not related to that variable. Further, recre-ations of these decompositions for the largest province in our sample, Ontario, yields very similar results, suggesting that this last marginal impact is not due to shifts in place of residence across cohorts. As a result, the observed difference appears to be due to shifts in the industrial distributions across entry cohorts. The Diffo line shows that even after accounting for all observable characteristic changes, there still remains a difference between the wage CDFs for the 1981 and 1998 entry cohorts of almost 20% through much of the wage range. Interestingly, the covariate composition changes affect higher wage earners more than lower earn-ers: the loss of access to union jobs and non-service sector jobs has the predictable impact of reducing access to higher wage (for workers of this education level) jobs. What is left after accounting for covariate shifts is a relatively even proportionate decline in the wage distribution, at least above the 20th percentile. The smaller declines below the 20th percentile may be due to the minimum wage supporting the bottom end of the distribution. We next repeat this exercise using the female sample and the estimates from the hazard based estimator applied to the female sample. In figure 4.14 we replicate the decomposition exercise reported in figure 4.13 for females. As for males, the overall difference between the 1981 and 1998 entry cohort CDFs takes a flattened U shape. For females, there is in fact no decline at the very bottom of the dis-tribution. The line portraying the impact of holding the minimum wage constant at its 1981 value shows that the minimum wage has a much more substantial on the female distribution, showing noticeable impacts as high as the 75th percentile. This is in line with results in DiNardo et al. (1996) showing that real minimum wage declines in the US have much larger impacts on female than male workers. As with the males, maintaining unionization rates their 1981 level reduces the 99 decline, particularly between the 25th and 75th percentile. However, the union effects are much smaller for females than males, generating a only 3% decline in the cross-cohort wage differential at the median for females but a 25% decline for males. The direct effect of changing union status to match the 1981 union struc-ture has virtually no effect on the wage distribution. In comparison, changes in the industrial structure are much more important for females than males. The over-all effect above the median of the wage distribution of holding minimum wages, unionization effects, and industrial composition constant is larger for females than males: leading to over a 50% decline in the cross-cohort wage differentials. As with males, larger covariate effects above the median lead to a flattening of the U-shaped cross-cohort differential. The line with all covariates held at their 1981 values and the minimum wage variables set to zero shows a relatively even decline across the wage distribution between the 1981 and 1998 entry cohorts. On aver-age that decline equals approximately 15%, which is quite similar to the observed decline for males. Thus, the smaller declines observed for females in the simple mean wage estimations appear to stem in large part from the role of the minimum wage. The minimum wage impact arises not because of dramatic changes in the minimum wage (the real decline from 1981 to 1998 is not dramatic relative to other changes facing low skilled workers) but because the minimum wage holds up the bottom part of the distribution, blocking larger decreases at the bottom end of the distribution. This effect is clearly visible in the fact that with all covariates, including the minimum wage, held at 1981 values, the 10th percentile of the dis-tribution would rise by 5% but with all covariates held constant and the minimum wage effect removed the 10th percentile declines by nearly 15%. The effect is much larger for females than males since more of the mass of the female wage distribution resides near the minimum wage. Thus, for both males and females a considerable portion of the cross-cohort decline in the wage distribution can be accounted for by declining access of new job entrants to unionized jobs in higher paying sectors. Nonetheless, even after accounting for these changes the 1998 job entry cohort faces a wage distribution that is shifted to the left by approximately 15% relative to the 1981 cohort, with that shift occurring relatively proportionately across the distribution. For females, this decline is hidden to some extent by the effect of minimum wages in forbidding wage declines at the very low end. 100 4.7 Conclusions In this paper, we investigate the underlying causes of large movements in cross-sectional age differentials in Canada over time. Following earlier work, we first conclude that these cross-sectional movements can best be described in terms of declines in wage levels across successive birth cohorts. Using rich data from a series of datasets, we then show that these declines can in turn be usefully summarized as stemming from declining wage levels across successive cohorts defined by their date of job entry rather than their date of birth. We find, further, that the wage-tenure profile changes little over time and that age differentials are relatively constant. An investigation of the job-cohort declines indicates that about half the decline can be accounted for by declines in unionization and changes in the industrial composition of job across cohorts. The minimum wage serves to reduce the effects of the general decline and these compositional changes at the lower end of the distribution. This is particularly true for women. In the end, we find evidence for a sizeable (on the order of 15% from the early 1980s to late 1990s cohorts) real decline in wages across cohorts. We also find that that decline happens roughly proportionately across the distribution. We argue that these patterns could be accounted for with a simple model in which rigidities associated with an implicit contract model convert known, long term trends into cross-cohort shifts in wage-tenure profiles. One implication of such a model is that we should observe cross-cohort shifts up in wage-tenure profiles in economies where real wages are trending up. This indeed appears to be the case in both the U K and Germany. The issue then becomes why one observes long term negative trends in wages for low skilled workers in some economies at the same time as long term positive trends in others. Beaudry and Green(2000b) provide a model explaining these trends in terms of an interaction of technological choice with differing rates of factor accumulation across economies. Indeed, they argue that their model can be used to explain the specific example of declining low skilled wages in the US and rising low skilled wages in Germany. Thus, our argument is that observed movements in age differentials can ultimately be related back to larger forces moving skill prices over time. We argue that this happens not because of a change in the price of experience in response to technological change but because of the way changes in the price of low educated labour interact with institutions in the labour market. 101 -i 1 1 1 r - i 1 r B - — Age 2 0 - 2 4 , 0 - 2 Yrs Tenure Age 2 5 - 3 4 , 0 - 2 Yrs Tenure Age 3 5 - 4 4 , 0 - 2 Yrs Tenure Age 2 5 - 3 4 , 9 - 1 0 Yrs Tenure Age 3 5 - 4 4 , 9 - 1 0 Yrs Tenure 1 9 8 1 1 9 8 3 1 9 8 5 1 9 8 7 1 9 8 9 1 9 9 1 1 9 9 3 1 9 9 5 1 9 9 7 1 9 9 9 Y e a r Figure 4.1: Mean Log Wages by Age Group, Year and Job Tenure, High School Men 102 i 1 1 r Age 2 0 - 2 4 , 0 - 2 Yrs Tenure Age 2 5 - 3 4 , 0 - 2 Yrs Tenure Age 3 5 - 4 4 , 0 - 2 Yrs Tenure Age 2 5 - 3 4 , 9 - 1 0 Yrs Tenure Age 4 5 - 5 4 , 9 - 1 0 Yrs Tenure J I I I I l_ 1 9 8 1 1 9 8 3 1 9 8 5 1 9 8 7 1 9 8 9 1 9 9 1 1 9 9 3 1 9 9 5 1 9 9 7 1 9 9 9 Y e a r Figure 4.2: Mean Log Wages by Age Group, Year and Job Tenure, High School Women 103 —i 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 r B- " A- - ' A 1981 Cohort, True A 1989 Cohort, True A 1998 Cohort, True B 1981 Cohort, Fitted B 1989 Cohort, Fitted • 1998 Cohort, Fitted 1 9 8 1 1 9 8 3 1 9 8 5 1 9 8 7 1 9 8 9 1 9 9 1 1 9 9 3 1 9 9 5 1 9 9 7 1 9 9 9 Year Figure 4.3: Mean Log Wages by Job Entry Cohort, High School Men 104 3 15 1 1 1 1 — I — — I 1 1 ~ i 1 1 1 1 1 1 1 3 10 - -3 05 A 1981 Cohort, True -3 00 A 1989 Cohort, True -2 95 -A 1998 Cohort, True o n n B 1981 Cohort, Fitted Z y U B 1989 Cohort, Fitted -2 85 O 1998 Cohort. Fitted -2 80 - _ CD cn 2 75 -D 5: 2 70 -C J I 2 65 -o _ i 2 60 - ~ ^ — c 2 55 D CD 2 50 - A " ~* 2 45 -2 40 -2 35 A- " 2 30 -2 25 _ • 2 20 - -2 15 -2 10 i I I i i i i i i i i 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 Year Figure 4.4: Mean Log Wages by Job Entry Cohort, High School Women 105 Figure 4.5: Effects of minimum wage on the wage distribution, High school men, 25-34 years old, manufacturing 106 Figure 4.6: Effects of union coverage on the wage distribution, High school men 25-34 years old, manufacturing 107 Figure 4.7: Effects of industry coverage rate on the wage distribution, High school men, 25-34 years old, manufacturing 108 o I 1 1 1 1 1 1 ' 1 1 r ci o — Services — -— Manufact. 6 o -I—» 6 0 10 2 0 30 Hour ly wage ( 1 9 9 8 do l l a r s ) Figure 4.8: Effects of industry on the wage distribution, High school men, 25-34 years old 109 -*-» '<n CD Q 10 20 Hour ly wage ( 1 9 9 8 do l l a r s ) Figure 4.9: Effects of entry cohort on the wage distribution, High school men, 25-34 years old, manufacturing 110 Figure 4.10: Effects of entry cohort on the wage distribution, no minimum wage, High school men, 25-34 years old, manufacturing 111 6 0 10 20 30 Hour ly wage ( 1 9 9 8 do l l a r s ) Figure 4.11: Effects of entry cohort on the wage distribution, , High school women, 25-34 years old, manufacturing 112 Figure 4.12: Difference between years, actual and fitted, high school men 113 Figure 4.13: Decomposition of Difference between Wage Distributions, High School Men, Ages 25-34, 1981 and 1998 114 Figure 4.14: Decomposition of Difference between Wage Distributions, High School Women, Ages 25-34, 1981 and 1998 115 Table 4.1: Differences in Log Wages B y Cohort and Year: Measured Relative to 25-34 Year Olds in 1981 1981 1989 1998 Males Cohort: 3 0.07(0.01) 0.09 (0.01) -0.02 (0.02) 4 0.00 0.06 (0.01) 0.06 (0.01) 5 - -0.08 (0.01) -0.01 (0.01) 6 - - -0.17 (0.01) Females Cohort: 3 0.02 (0.01) -0.03 (0.01) 0.03 (0.02) 4 0.00 -0.01 (0.01) 0.07 (0.01) 5 - -0.07 (0.01) 0.03 (0.01) 6 - - -0.08 (0.01) Note: Cohort 3 consists of individuals who were 35 to 44 in 1981. Cohort 4 consists of individuals who were 25 to 34 in 1981. Cohort 5 consists of individuals who were 25 to 34 in 1989. Cohort 6 consists of individuals who were 25 to 34 in 1998. We follow cohort 3 by examining wages of individuals who were 45 to 54 in 1989 and 55 to 64 in 1998. Other cohorts are followed in similar fashion. The matchings across years for a cohort are clearly only approximate. A l l entries in the table represent the difference between the average log wage in a given year for a given cohort relative to cohort 4 in 1981. The data pertain to individuals with a high school or less education. 116 Table 4.2: Mean Log Wage Regressions, 1981, 1989 and 1998 Data Variable Constant Age Dummies Tenure Tenure Sq'd Males 2.44 (.010)* 16-19 -0.50 (.010)* 20-24 -0.19 (.010)* 35-44 0.09 (.010)* 45-54 0.12 (.014)* 55-64 0.043 (.027) 0.014 (.0025)* 0.000013 (.00016) Females 2.24 (.011)* -0.30 (.011)* -0.12 (.012)* 0.0095 (.010) -0.0070 (.013) 0.020 (.029) 0.032 (.0027)* -0.00064 (.00016)* Job Entry Cohort Dummies 1980/81 1984/85 1988/89 1992/94 0.20 (.010)* 0.16 (.019)* 0.11 (.010)* 0.09 (.020)* 0.13 (.011)* 0.031 (.019) 0.068 (.011)* 0.072 (.021)* Age-Tenure Interactions # of observations 16-19 -.023 (.006)* 20-24 .0040 (.0034) 3 5 . 4 4 ..0056 (.0015)* 45-54 -.0079 (.0019)* 55-64 -.0089 (.0033)* 26463 -.014 (.0074)+ .0088 (.0045)* -.00011 (.0017) -.0026 (.0020) -.0082 (.0033)* 22058 Notes: Base age category is 25 to 34 year olds and the base cohort category is 1998 new job entrants. The age-tenure interaction terms are coefficients of the linear years of job tenure variable with the specified age group dummy variable. * indicates coefficient is statistically significantly different from zero at 1% level of significance. + indicates statistical significance at 5% level of significance. 117 Table 4.3: Covariate Distributions for 1981, 1989 and 1998 Job Entrant Cohorts 1981 Job Cohort 1989 Job Cohort 1998 Job Cohort Variables 1981 1989 1998 1989 1998 1998 Males Industry 0.09 Primary 0.11 0.06 0.04 0.07 0.05 Construction 0.18 0.07 0.02 0.20 0.08 0.15 Manufacturing 0.25 0.36 0.33 0.24 0.31 0.25 Transportation & 0.06 Communications 0.11 0.12 0.23 0.11 0.16 Trade 0.16 0.19 0.11 0.19 0.16 0.20 Financial Services 0.02 0.01 0.06 0.01 0.04 0.02 Health & Educ. 0.03 0.03 0.11 0.03 0.07 0.02 Other Services 0.13 0.09 0.02 0.10 0.10 0.20 Public Admin. 0.02 0.06 0.07 0.04 0.04 0.02 Union Coverage 0.41 0.44 0.63 0.35 0.50 0.17 Coverage Rate 0.25 in Industry 0.38 0.39 0.42 0.38 0.33 Min. Wage (1998$) 6.78 5.93 6.50 5.92 6.63 6.46 Females Industry 0.02 Primary 0.02 0.01 0.02 0.03 0.01 Construction 0.02 0.01 0.01 0.02 0.01 0.02 Manufacturing 0.19 0.11 0.14 0.17 0.16 0.18 Transportation & 0.02 Communications 0.04 0.06 0.12 0.05 0.08 Trade 0.19 0.21 0.20 0.24 0.18 0.26 Financial Services 0.10 0.07 0.08 0.08 0.07 0.05 Health & Educ. 0.11 0.17 0.27 0.11 0.21 0.10 Other Services 0.27 0.25 0.05 0.25 0.21 0.34 Public Admin 0.06 0.11 0.10 0.05 0.08 0.01 Union Coverage 0.22 0.40 0.52 0.26 0.46 0.11 Coverage Rate In Industry 0.31 0.38 0.40 0.31 0.34 0.21 Min. Wage (1998$) 6.74 5.94 6.52 5.92 6.60 6.50 118 Table 4.4: Job En t ry Cohort Effects from M e a n Log Wage Regressions High School Males Specification 1 2 3 Job Entry Cohort Nonunion Union 1980/81 0.22 (.01)* 0.17 (.01)* 0.15 (.01)* . 21 (.02)* 1982/83 0.19 (.01)* 0.17 (.01) 0.13 (.01) .24 (.02)* 1984/85 0.17 (.01)* 0.03 (.01)* 0.09 (.01)* .20 (.02)* 1986/87 0.14 (.01)* 0.11 (.01)* 0.08 (.01) .17 (.02)* 1988/89 0.11 (.01)* 0.07 (.01)* 0.04 (.01)* . 15 (.02)* 1990/91 0.09 (.01)* 0.07 (.01)* 0.04 (.01)* .13 (.02)* 1992/94 0.10 (.01)* 0.09 (.01)* 0.08 (.01)* .09 (.02)* 1995/96 0.06 (.01)* 0.06 (.01)* 0.06 (.01)* .07 (.02)* Other Controls: Unemployment Rate yes yes yes Industry Indicators no yes ; yes Province Indicators no yes yes Union Status no yes yes Union Rate in Industry no yes yes Union Status Interactions with: Industry no no yes Province no no yes Union Rate in Industry no no yes # of observations 68987 68987 68987 Adjusted R-squared 0.217 0.372 0.387 Notes: The base cohort category is 1998 new job entrants. * indicates coefficient is statistically significantly different from zero at 1% level of significance. + indicates statistical significance at 5% level of significance. 119 Table 4.5: Job En t ry Cohort Effects from M e a n Log Wage Regressions High School Females Specification 1 2 Job Entry Cohort Nonunion Union 1980/81 1982/83 1984/85 1986/87 1988/89 1990/91 1992/94 1995/96 Other Controls: Unemployment Rate Industry Indicators Province Indicators Union Status Union Rate in Industry Union Status Interactions with: Industry Province Union Rate in Industry # of observations Adjusted R-squared 0.11 (.01)* 0.07 (.01)* 0.07 (.01)* 0.04 (.01)* 0.07 (.01)* 0.05 (.01)* 0.06 (.01)* 0.04 (.01)* yes no no no no no no no 58762 0.158 0.05 (.01)* 0.01 (.01) 0.05 (.01)* 0.05 (.01)* 0.02 (.01)* 0.04 (.01)* 0.05 (.01)* 0.03 (.01)* yes yes yes yes yes no no no 58762 0.300 0.04 (.01)* -0.002 (.01) 0.02 (.01)* 0.005 (.01) 0.02 (.01)* 0.01 (.01) 0.04 (.01)* 0.03 (.01)* .13 (.02)* .11 (.02)* .14 (.02)* .12 (.02)* .14 (.02)* .15 (.02)* .11 (.02)* .03 (.02) yes yes yes yes yes yes yes yes 58762 0.308 Notes: The base cohort category is 1998 new job entrants. * indicates coefficient is statistically significantly different from zero at 1% level of significance. + indicates statistical significance at 5% level of significance. 120 Appendix A Auxiliary regressions 121 Table A . l : Mean Log Wage Regressions: Males Additional Specifica-tions Using Data From AH Years Variable Simple Regression Including Industry and Allowing Union Interations Constant Union Coverag 2.52(.02)* 2.32(.02)* 0.21*(.01)* 2.24(.02)* 0.32(.02)* Age Dummies 16-19 20-24 35-44 45-54 55-64 -0.46(.01)* -0.02(.01)* 0.26(.01)* 0.31(.01)* 0.20(.02)* -0.33(.01)* -0.03(.01)* 0.22(.01)* 0.29(.01)* 0.23(.01)* -0.30(.01)* -O.Ol(.Ol)* 0.25(.01)* 0.30(.01)* 0.21(.01)* Age Dummies*Coverage 16-19 - - -0.06(.04) 20-24 - - -0.09(.01)* 35-44 - - -0.09(.01)* 45-54 - - -0.08(.02)* 55-64 - - 0.08(.04)* Tenure 0.041(.001)* 0.034(.001)* 0.044(.002)* Tenure Sq'd -.0007(.0001)* -.0008(.0001)* -.OOl(.OOOl)* Tenure*Coverage - - -.30(.002)* Tenure Sq'd*Coverage - - .OOll(.OOOl)* Industry Coverage Rate - .46(.01)* .61(.02)* Industry Rate * Coverage - - -.24(.03)* Unemployment Rate -.03(.003)* -.03(.002)* -.03(.002)* No. of observations 68987 68987 68987 Adjusted R-squared .216 .372 .387 Notes: Base age category is 25 to 34 year olds and the base cohort category is 1998 new job entrants. A l l columns include age-tenure interactions, with column 3 including these interactions times union status. Columns 2 includes industry dummy variables corresponding to the two digit industry level. Column 3 includes these dummies and their interaction with coverage. * indicates coefficient is statistically significantly different from zero at 1% level of significance. **indicates statistical significance at 5% level of significance. 122 Table A.2: Mean Log Wage Regressions: Males Additional Specifica-tions Using Data From All Years Variable Simple Regression Including Industry and Allowing Union Interations Constant Union Coverage 2.20(.03)* 2.04(.03)* 0.18*(.01)* 2.05(.03)* 0.03(.03) Age Dummies 16-19 20-24 35-44 45-54 55-64 -0.29(.01)* -0.02(.01)* O.ll(.Ol)* 0.07(.01)* 0.08(.02)* -0.24(.01)* -0.03(.01)* O.lO(.Ol)* 0.08(.01)* 0.12(.02)* -0.22(.01)* -0.02(.01)* O.ll(.Ol)* 0.08(.01)* 0.12(.02)* Age Dummies*Coverage 16-19 - - -0.11(.05) 20-24 - - -0.06(.02)* 35-44 - - -0.06(.01)* 45-54 - - -0.02(.02)* 55-64 - - 0.01(.05)* Tenure 0.049(.001)* 0.034(.001)* 0.039(.002)* Tenure Sq'd -.OOll(.OOOl)* -.0008(.0001)* -.0009(.0001)* Tenure*Coverage - - -.021(.003)* Tenure Sq'd*Coverage - - .0008(.0002)* Industry Coverage Rate - .56(.02)* .57(.03)* Industry Rate * Coverage - - .08(.05)** Unemployment Rate -.01(.002)* -.01(.003)* -.01(.003)* No. of observations 68987 68987 68987 Adjusted R-squared .216 .372 .387 Notes: Base age category is 25 to 34 year olds and the base cohort category is 1998 new job entrants. A l l columns include age-tenure interactions, with column 3 including these interactions times union status. Columns 2 includes industry dummy variables corresponding to the two digit industry level. Column 3 includes these dummies and their interaction with coverage. * indicates coefficient is statistically significantly different from zero at 1% level of significance. **indicates statistical significance at 5% level of significance. 123 Bibliography Acemoglu, Daron and Jrn-Steffen Pischke, "The Structure of Wages and Investment in General Training," Journal of Political Economy, 1999, 107 (3), 539-72. Baggs, Jennifer J . , "Trade Barriers, International Business, and the Environ-ment." PhD dissertation, University of British Columbia 2002. Baldwin, John R., Desmond Beckstead, and Richard Caves, "Changes in the Diversification of Canadian Manufacturing Firms (1973-1997): A Move to Specialization," 2001. Statistics Canada, Analytical Studies Research Paper No 179. Beaudry, Paul and David Green, "Cohort patterns in Canadian earnings: assessing the role of skill premia in inequality trends," Canadian Journal of Economics, November 2000, 33 (4), 907-906. _ and John Dinardo, "The Effect of Implicit Contracts on the Movement of Wages Over the Business Cycle: Evidence from Micro-Data," Journal of Political Economy, 1991, 99, 665-688. Beaulieu, Eugene, "The Canada-U.S. Free Trade Agreement and labour market adjustments in Canada," Canadian Journal of Economics, 2000, 33, 540-63. Brander, James and Barbara Spencer, "Unionized Oligopoly and Interna-tional Trade Policy," Journal of International Economics, 1988, 24, 217-34. Buchinsky, Moshe, "Changes in the U.S. Wage Structure 1963-1987: Application of Quantile Regression," Econometrica, March 1994, 62 (2), 405-58. Campbell, Bruce, Maria Teresa Gutierrez Haces, Andrew Jackson, and Mehrene Larudee, Pulling Apart: The Deterioration of Employment and In-124 come in North America under Free Trade, Canadian Centre for Policy Alterna-tives, 1999. Cox, D. and Rick Harris, "A Quantitative Assessment of the Economic Impact of Sectoral Free Trade with the United States," Canadian Journal of Economics, 1984, 19, 377-94. Dinardo, John and Thomas Lemieux, "Diverging Male Wage Inequality in the United States and Canada, 1981-88: Do Institutions Explain the Difference?," Industrial and Labour Relations Review, July 1997, 50 (4), 629-51. DiNardo, John, Nicole Fortin, and Thomas Lemieux, "Labour Market Institutions and the Distribution of Wages, 1973-1992: A Semiparametric Ap-proach," Econometrica, September 1996, 64 (5), 1001-44. Donald, Stephen G. , David Green, and Harry J. Paarsch, "Differences in Wage Distributions between Canada and the United States: An Application of a Flexible Estimator of Distribution Functions in the Presence of Covariates," Review of Economic Studies, October 2000, £7(4), 609-33. Farber, H . and D. Saks, "Why workers want unions: the role of relative wages and job characteristics," Journal of Political Economy, 1980, 88, 349-69. Freeman, Richard and J.Medoff, What Do Unions Do?, New York, NY: Basic Books, 1984. _ and Lawrence Katz, "Industrial Wage and Employment Determination in an Open Economy," in John Abowd and Richard Freeman, eds., Immigration, Trade, and the Labour Market, Chicago: University of Chicago Press, 1991. Gaston, Noel and Daniel Trefler, "Union wage sensitivity to trade and pro-tection: Theory and evidence," Journal of International Economics, 1995, 35, 1-25. _ and _ , "The labour market consequences of the Canada-U.S. Free Trade Agreement," Canadian Journal of Economics, 1997, 30, 18-41. Gera, S. and G. Grenier, "Interindustry wage differentials and efficiency wages: some Canadian evidence," Canadian Journal of Economics, 1994, 27, 81-100. 125 Gosling, Amanda, Stephen Machin, and Costas Meghir, "The Changing Distribution of Male Wages in the UK," Review of Economic Studies, Ocotober 2000, 67(4), 635-66. Green, David and James Townsend, "An Investigation into the Sources of Declining Youth Wages in Canada," presented at the 34th Annual Meeting of the Canadian Economics Association, Vancouver, BC, June 2000. Grossman, Gene, "Partially Mobile Capital: A general approach to two-sector trade theory," Journal of International Economics, 1983, 15, 1-17. Grout, P., "Investment and Wages in the Absence of a Binding Contract: A Nash Bargaining Approach," Econometrica, 1984, 52, 449-60. Gu, Wulong, Gary Sawchuk, and Lori Whewell, "The Effect of Tariff Re-ductions on Firm Size and Firm Turnover in Canadian Manufacturing," 2002. Statistics Canada Research Paper. Hazledine, T., "Why Do the Free Trade Gain Numbers Differ So Much?," Cana-dian Journal of Economics, 1990, 23, 791-806. Head, Keith and John Ries, "Rationalization Effects of Tariff Reductions," Journal of International Economics, 1999, 47, 295-320. Johnson, G., "Work Rules, Featherbedding, and Pareto-optimal Union Manage-ment Bargaining," Journal of Labor Economics, 1990, 8, 237-259. Johnson, Susan, "The Canadian Union Density 1980-1998 and Prospects for the Future: An Empirical Investigation," 2001. Wilfred Laurier University Working Paper. Juhn, Chinhui, Kevin M . Murphy, and Brooks Pierce, "Wage Inequality and the Rise in Returns to Skill," Journal of Political Economy, June 1993, 101 (3), 410-42. Kapsalis, C. , R.Morissette, and Garnett Picot, "The Returns to Education, and the Increasing Wage Gap Between Younger and Older Workers," Statistics Canada Analytical Studies Branch Research Paper no. 131 1999. 126 Katz, Lawrence and Lawrence Summers, "Industry Rents: Evidence and Implications," Brookings Papers on Economic Activity: Microeconomics, 1989, pp. 209-75. Kuhn, Peter, "Unions and the economy: What we know; What we should know," Canadian Journal of Economics, 1998, 31, 1033-56. Lemieux, Thomas, "Unions and wage inequality in Canada and the United States," in David Card and richard Freeman, eds., Small Differences that Matter, Chicago: University of Chicago Press, 1993. _ , "Estimating the Effects of Unions on Wage Inequality in a Panel Data Model with Comparative Advantage and Nonrandom Selection," Journal of Labor Eco-nomics, April 1998, 16 (2), 261-91. _ and Nicole Fortin, "Public Policies in a Labour Market in Transition," in W.C. Riddell and F. St. Hilaire, eds., Income Redistribution in Canada: Mini-mum Wages versus Other Policy Instruments, IRPP, 2000. Lewis, H.G. , Unionism and Relative Wages in the United States: An Empirical Inquiry, Chicago: University of Chicago Press, 1963. MaCurdy, Thomas and Thomas Mroz, "Measuring Macroeconomic Shifts in Wages from Cohort Specifications," 1995. mimeo Stanford University. Magun, S., S. Rao, B. Lodh, L. Lavalee, and J. Pierce, "Open Borders: An Assessment of the Canada-U.S. Free Trade Agreement," Economic Council of Canada Discussion Paper No.344 1988. Mezzetti, C. and E . Dinopoulos, "Domestic unionization and import compe-tition," Journal of International Economics, 1991, 31, 79-100. Morissette, Rene, "The declining labour market status of young men," Statistics Canada Income Research Paper Series 1997. _ , "The growth of earnings inequality in Canada," Statistics Canada Income Research Paper Series 1997. Schmitt, John, "The Changing Structure of Male Earnings in Britain, 1974-1988," in Richard Freedman and Lawrence Katz, eds., Changes and Differences in Wage Structures, Chicago: University of Chicago Press, 1995. 127 Trefler, Daniel, "The Long and Short of the Canada-U.S. Free Trade Agree-ment," 2001. National Bureau of Economics Working Paper. 128 


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items