The sources of interindustry wage differentials

8 ISER Working Paper Series The Sources of Interindustry Wage Differentials Priscila Ferreira Institute for Social and Economic Research, University of Essex Department of Economics, University of Minho No. 2009-13 March 2009 www.iser.essex.ac.uk Non-technical summary Empirical analysis using cross-sectional data commonly …nd signi…cant wage di¤erences across industries. That is, some industries appear to pay higher wages than others to what appear to be equal workers employed in what appear to be equal …rms. However, in a perfectly competitive labour market with similar workers and …rms, these di¤erences should not exist as wage di¤erences across industries will encourage workers to move between …rms, which would equalises wages. To the extent that inter-industry wage di¤erences may re‡ect the existence of non-competitive mechanisms, they are an empirical …nding di¢cult to explain in labour economics. In this paper we investigate the existence and sources of inter-industry wage di¤erences. First, we analyse the size and persistence of wage di¤erences between industries. Second, we look at the e¤ect of the characteristics of the worker, the …rm and the match between the two. This allows us to answer questions related to (i) the nature of wage di¤erences: is it because wages are truly di¤erent, or are they due to di¤erent types of workers employed by the industries, and (ii) which of these e¤ects are the most important in explaining di¤erences in wages across industries? Analysis of the period 1986-2000 suggest the existence of important inter-industry wage di¤erentials. Workers with the same characteristics working in …rms with observably equal characteristics have di¤erent wages depending on the industry in which they are employed. The wage di¤erences identi…ed are considerable, with some industries paying wages that are more than 41% above the economy average while others pay wages that are more than 26% below the average. Amongst the industries paying the lowest wages, we …nd the Manufacture of furniture, Textiles, Clothing and Restaurants and cafés. The industries paying the highest wages tend to be related to …nancial intermediation like Banking or Insurance, the Electricity, gas and water supply services. However, the estimated inter-industry wage structure is greatly weakened when controlling for unobserved characteristics of workers (e.g., unobserved ability), suggesting that the raw di¤erences are due to the concentration of high wage workers in certain industries and not to genuine di¤erences in wages paid by …rms across industries. Further analysis, controlling for unobserved characteristics of …rms, shows that compensation policies of …rms vary across industries and that by changing industries workers may enjoy substantial wage growth. We thus conclude that inter-industry wage di¤erences re‡ect di¤erent treats given by …rms across industries. Why might …rms treat their workers by paying them wages above the economy average? Our results indicate that workers who are paid more than the economy average are less likely to quit. Therefore, we conclude that there is a possibility that …rms gain from long term employment relationships. In these circumstances above-average wages are a pro…t maximising strategy because the costs of higher wages may be o¤set by the bene…ts of reduced turnover. This is consistent with a labour market in which industry speci…c skills are important and where e¢ciencies are gained from creating incentives to worker-…rm attachments. The sources of interindustry wage di¤erentials. Priscila Ferreiray University of Essex, U.K. and University of Minho, Portugal 31st March 2009 Abstract We analyse the nature of interindustry wage di¤erentials using Portuguese data. Estimates from models controlling for observed worker and …rm characteristics reveal signi…cant and persistent raw interindustry di¤erentials, which questions the competitive model of the labour market. However, estimates controlling for unobserved worker heterogeneity suggest that the raw di¤erentials are due to the concentration of high wage workers in certain industries and not to genuine di¤erences in compensation across industries. However, a complete decomposition shows that (i) …rm e¤ects on average explain 70% of the industry wage premia, and (ii) genuine and sizeable interindustry wage di¤erentials exist. These di¤erentials are shown to increase the time to separation from …rms, and are therefore compatible with the competitive model. Keywords: interindustry wage di¤erentials; unobserved worker …rm and match e¤ects; job mobility JEL Classi…cation: C33, J21, J31, J62, J63. I thank Stephen Jenkins for discussions. I also thank Mark Taylor, and participants in the Warsaw International Economic Meeting 2008, the Leibniz Seminar on Labor Research (Berlin Network of Labor Market Research, November 2008), the Brown Bag Seminar at the Toulouse School of Economics (March 2009), and the 8th GEP PG Conference for helpful comments and suggestions. I am grateful to the Statistics Department, Ministry of Employment, Portugal, for access to Quadros de Pessoal. Funding by Fundação para a Ciência e a Tecnologia (contract SFRH/BD/14713/2004) is also acknowledged. y Correspondence to: ISER, University of Essex, Colchester CO4 3SQ, UK. E-mail: [email protected] 1 Introduction Empirical analyses using cross-sectional data commonly …nd signi…cant wage di¤erentials across industries. That is, some industries appear to pay higher wages than others to observably equal workers employed in observably equal …rms. However, in a perfectly competitive labour market with homogeneous workers and …rms, these di¤erences should not exist as wage di¤erences across industries will prompt mobility of workers that equalises wages. To the extent that interindustry wage di¤erentials may re‡ect the existence of noncompetitive mechanisms, they are an empirical …nding di¢cult to explain in labour economics. Noncompetitive mechanisms, however, are not the only possible explanation for these di¤erentials. Such a wage structure can arise because our observed measures of the qualities of workers are imperfect. Therefore, high wage industries may either be hiring high wage workers or be composed of high wage …rms, or have a combination of both. This means that substantial wage di¤erences across industries can be either a consequence of the type of workers employed, or of di¤erent compensation policies of …rms. A …rst purpose of this paper is to investigate the existence and sources of interindustry wage di¤erentials. Are cross-sectional industry wage di¤erentials true wage di¤erences, or are they a consequence of unobserved individual heterogeneity? What is the relative importance of unobserved worker, …rm and match e¤ects in explaining di¤erences in wages across industries? Due to lack of appropriate longitudinal linked employer and employee data, decompositions of wage di¤erentials are not common in the literature. The Portuguese data set used here, Quadros de Pessoal, enables us to estimate worker, …rm and match unobserved e¤ects using the techniques developed in Abowd et al. [AKM] (1999) and Woodcock (2008a), and to decompose the interindustry wage di¤erentials into proportions attributable to each of these unmeasured components. Therefore, our analysis contributes to the ongoing debate on the sources of the industry wage structure, and adds to existing evidence for France and the U.S.A. If interindustry wage di¤erentials exist and persist over time, even after controlling for worker and …rm characteristics, then wages play roles other than providing signals for labour reallocation. That is, wages do not re‡ect temporary di¤erences in productivity caused by Comments on earlier versions of this paper were received at presentations in the Warsaw International Economic Meeting 2008, the Leibniz Seminar on Labor Research (Berlin Network of Labor Market Research, November 2008), and the Brown Bag Seminar at the Toulouse School of Economics (March 2009). 1 shifts in the relative demand for labour between industries. There are two possible explanations for the existence of wage di¤erences across industries. Either …rms are not maximising pro…ts, or …rms …nd it pro…table to pay higher wages. The latter is the main hypothesis of e¢ciency wage models, which argue that higher wages can increase output and so wages above opportunity costs are pro…t maximising. Although e¢ciency wage models can be grouped in categories such as shirking, turnover, adverse selection and fair wages, these categories are not mutually exclusive. Firms may be paying higher wages to accomplish a combination of objectives. However, we can use them individually to test the competitive model. One possibility is to analyse the relationship between the industry wage premia and turnover. Krueger and Summers (1988) suggest that if workers in high wage industries truly receive economic rents then we could expect to …nd a negative relationship between turnover and interindustry wage di¤erentials. In this case the cost of higher wages would be at least partially o¤set by the bene…ts derived from reduced turnover rates. In this context, the second aim of this paper is to investigate the relationship between the industry wage premia and separations from …rms. This analysis is done using duration models in which the dependent variable is the time to separation from …rms, and where the industry wage premia is included amongst the explanatory variables. Our main …ndings can be summarised as follows. Firstly, we …nd that wages di¤er across industries in Portugal and that the dispersion of wage di¤erentials is fairly stable over time. Hence, temporary variations in productivity are not the main force driving the industry wage structure. Secondly, our decomposition of the raw interindustry wage di¤erentials into components due to worker, …rm and match e¤ects reveals that unobserved …rm e¤ects are the major source of the observed di¤erences in wages. Therefore, interindustry job mobility can have a large impact on the wages received as the nature of the di¤erences is not due to a portable component of compensation (worker e¤ects), but to di¤erent compensation policies across …rms. Thirdly, we …nd that the industry wage premium is positively associated with the time to separation from …rms. This suggests that the mechanisms generating wage dispersion across industries may be compatible with predictions of the competitive model, insofar as it can be pro…t maximising for …rms to pay higher wages in order to reduce turnover costs. The paper is structured as follows. The next section discusses the issues associated to the identi…cation of interindustry wage di¤erentials in empirical analyses. In Section 4, we docu2 ment cross-sectional di¤erences in wages across industries in Portugal and identify their sources. Section 5 describes the statistical approach used to identify the importance of unobserved heterogeneity in generating these di¤erentials. In Section 6 we decompose the raw interindustry wage di¤erentials into proportions due to unobserved person, …rm and match e¤ects. In Section 7, the competitive model is tested by analysing the relationship between the time workers take to separate from …rms and the industry wage premia identi…ed in the previous section. Summary and conclusions are presented in Section 8. 2 Theoretical background In a competitive model where all workers and jobs are homogeneous, information and search costs are low, and issues of worker motivation and risk shifting are not important, the long run labour market equilibrium is characterized by identical wages for all workers and little unemployment. Transitory wage di¤erentials that re‡ect di¤erentials in labour productivity (caused by demand ‡uctuations at the …rm level, or at any labour market segment level) induce labour mobility. As there is no reason for workers to form attachments to speci…c …rms or industries, they will respond to wage di¤erentials and move between segments, equalizing productivity and restoring wage equality across segments. Therefore, in equilibrium, the competitive model with perfect information entails all workers with the same worker/job match characteristics obtaining the same wage. The model implies the nonexistence of unemployment, because wages adjust until the demand for workers equals their supply and the labour market clears. This suggests that wage dispersion and unemployment are closely related. Di¤erences in wages generate worker ‡ows across …rms, and transitions between jobs can involve a period of unemployment. The model also predicts the nonexistence of wage di¤erentials associated with the industry where the worker is employed, except if these are compensating di¤erentials for nonpecuniary aspects of the job. However, reality appears not to comply with these predictions, as introducing the industry of employment commonly adds explanatory power to a wage equation which explains wages solely in terms of worker, …rm and worker/…rm match characteristics. The existence and persistence of wage di¤erentials leads us to the question of how wages are determined. This is an important issue, because understanding the wage determination process is fundamental to 3 understanding labour mobility and unemployment. Responses in the literature to observed interindustry wage di¤erentials range from denial of their existence to accepting that these di¤erences are true. The …rst approach complies with the competitive model and argues that observed cross-sectional wage di¤erences between industries are illusory rather than re‡ecting true industry di¤erences in compensation. According to this view the observed wage structure is generated by unobserved heterogeneity of workers and job characteristics. Higher wage industries may be compensating workers for their unmeasured labour quality, or for some less desirable working conditions or job characteristics that a¤ect the utility of workers. If measures of the worker’s productive abilities are imperfect and if workers in high wage industries have more productive ability than others, then the industry wage premia could simply be re‡ecting the earnings capacity of its workforce. In this case, changing industries will not be associated with wage gains or losses for workers because the observed di¤erences are due to a portable component of compensation that will follow the worker to whatever industry he moves to. The second approach accepts that there are interindustry wage di¤erentials even when controlling for the nature of work and the quality of workers. One possible cause of such wage di¤erentials is the existence of worker-…rm attachments. These appear if e¢ciencies are gained when workers remain with speci…c …rms or industries for extended periods. If e¢ciencies outweigh the gains in productivity that might come from reallocating labour in response to every transitory demand shock, then the labour market structure, employment rules, and wage structures adjust to encourage long-term attachments and to limit day-to-day competition in the labour market. The result is a segmented labour market where market competition is limited, and workers become attached to speci…c …rms or industries with competition taking place only at speci…c ports of entry to internal labour markets. If workers do not compete in a single aggregate labour market, the labour market is expected to adjust less rapidly than suggested by the neoclassical model because wages play roles other than providing signals for labour reallocation. E¢ciency wage models provide some alternative explanations for the existence of industry speci…c wages. Krueger and Summers (1988) and Thaler (1989) group these models into four categories: (i) in shirking models, high wage industries should be those with high monitoring 4 costs and that have relatively higher costs of employee shirking; (ii) in turnover models, high wage industries are those in which turnover costs are highest; (iii) in adverse selection models, high wage industries are those more sensitive to labour quality di¤erences or have higher costs of measuring quality; (iv) in fair wage models, industries with high pro…ts will pay higher wages, because workers believe that fairness requires …rms to share rents. Therefore, despite predictions from the competitive model that a pro…t maximising …rm o¤ers wages equal to the value of marginal productivity of labour, there can be reasons why …rms pay supra competitive wages and create incentives for long term attachments of its workforce. These can be part of a pro…t maximising strategy of …rms, and hence not fully incompatible with the neoclassical model. When longitudinal data on workers became available, within-worker transformations and …rst di¤erenced regressions were commonly applied in empirical studies of interindustry wage di¤erentials, in the 1980s and early 1990s, to eliminate the e¤ect of worker unobserved heterogeneity, and to identify the sources of the industry e¤ects estimated using cross-sectional data. The results are as varied as the predictions of the models discussed. Murphy and Topel (1987) …nd evidence that two thirds of the observed di¤erential can be explained by unmeasured worker characteristics. On the other hand, despite not identifying the sources of interindustry wage di¤erentials, Krueger and Summers (1987) …nd empirical regularities that lead them to conclude that unmeasured worker characteristics cannot explain such di¤erentials. These regularities include evidence suggesting that: (i) by changing industries workers receive wage changes similar to the industry e¤ects found in cross-sectional data; (ii) the industry premia/penalty is similar for di¤erent types of quality of workers; and (iii) wage di¤erentials can be explained by product market characteristics. Blackburn and Neumark (1992), by including measures of worker abilities (test scores) in their regressions, conclude that ability can only account for a small portion (10%) of interindustry wage di¤erences observed in cross-sectional analysis. Gibbons and Katz (1992) conclude that a major proportion of interindustry wage di¤erences cannot be explained by the sorting of workers across industries by unobserved productive ability. More recently, questions of the sources of the industry wage structure have resurfaced due to the emergence of longitudinal linked employer-employee data. Yet, the results using this type of data remain varied. Using French data, Goux and Maurin (1999) …nd that the wage structure 5 is mainly due to unmeasured labour quality and that the potential wage gain from switching industries would be less than 3%. Furthermore, the authors conclude that these remaining true di¤erentials do not persist over time. Also using French data, AKM (1999) conclude that person e¤ects are relatively more important in explaining the di¤erentials found in crosssectional analysis. The same result is obtained by Abowd, Finer and Kramarz (1999) with data for the State of Washington and applying the same decomposition as AKM (1999). Woodcock (2008) using American data …nds that, controlling for match e¤ects, …rm e¤ects are responsible for 72% of the variance in raw interindustry wage di¤erentials.1 The di¤erence between the …ndings of Goux and Maurin (1999) and other papers using matched worker-…rm data might be due to the di¤erent techniques applied. The other studies use similar statistical methods to decompose the raw interindustry wage di¤erentials found in cross-sectional data, and all provide evidence of the existence of pure interindustry wage di¤erentials. This paper will add to the small empirical literature that controls for measured and unmeasured characteristics of both sides of the labour market. The methodology developed in AKM (1999) is used to distinguish between the two leading explanations of the industry wage structure found in cross-sectional analysis. 3 Data used and the Portuguese labour market 3.1 The Quadros de Pessoal data The data used in this analysis is the Quadros de Pessoal (Lists of Personnel) from Portugal. The Quadros de Pessoal is a longitudinal data set with matched information on workers and …rms. Since 1985, the survey has been annually collected (in March until 1993, and in October from 1994 onwards) by the Portuguese Ministry of Employment and the participation of …rms with registered employees is compulsory. The data include all …rms (about 200 thousand per year) and employees (about two million per year) within the Portuguese private sector. The analyses in this paper are derived from data collections for each year from 1986 to 2000, with 1990 excluded because the database was not built in that year. Although the survey continues, 1 In contrast with the other studies using linked employer-employee that disaggregate industry coding to a detailed level (more than 90 industries), Woodcock (2008) uses only 8 SIC Major Divisions. 6 the data currently available for analysis ends in 2000. Each …rm and each worker has a unique registration number which allows them to be traced over time. All information on both …rms and workers is reported by the …rm. In general, the information refers to the situation observed in the month when the survey is collected. In some cases, namely information on dates, reported data may refer to dates in the past (i.e., before the data collection month or to previous years) but is limited to the past within the speci…c …rm where the worker is employed. Information on workers includes, for example, gender, age, education level, level of skill, occupation, date of admission in the …rm, date of last promotion, monthly wages (split into some of its components) and monthly hours of work. Firm level data include, for example, the industry, location, number of workers, number of establishments, and legal structure. Some data management was carried out before implementing any analysis. First, we converted the data from a set of time series-cross sections into longitudinal panel data format. Second, to overcome computer memory size limitations, a 10% random sample of workers was selected from the cleaned panel data set. Third, because we will compute estimates of unobserved person and …rm e¤ects, we select only the observations that belong to the largest group of connected workers and …rms. This sample contains 1,823,572 observations related to 377,866 workers, 98,438 …rms and 589,826 matches over time.2 3.2 The Portuguese labour market The labour market in Portugal is very regulated. Employment protection regulation covers a wide range of issues such as the conditions under which individual dismissal are fair or unfair, the procedures for individual (and collective) notice and dismissal, severance payments, rules to the use of …xed-term contracts and of temporary work agency employment. In general, labour market law favours employment security and gives preference to contracts of employment of inde…nite duration, and most collective agreements (agreements reached after collective bargaining) make provision for career paths within the …rm. This is achieved by, 2 More details on the sample selection and construction of groups of connected workers and …rms can be found in Ferreira, (2009b). 7 e.g. de…ning a set of general rules on the criteria for automatic and merit-based promotions.3 However, while a promotion (career progress) can be decided unilaterally and requires only an agreement between the worker and the …rm, downgrading a worker to a lower category is more di¢cult. Downgrading can not be decided unilaterally, and even when the worker and the …rm agree with it a special justi…cation and authorization from the Ministry of Employment is required. Wages are negotiated yearly and are updated in January (if collective bargaining takes longer to reach an agreement, wages are updated retroactively). Collective negotiations usually occur at the occupation and industry level, and coverage of workers is generally irrespective of union membership. This is a result of existing mechanisms of extension of contracts: (i) an employer accepting an agreement usually applies it to all of its workforce; and (ii) under certain circumstances the Government can extend the negotiated contracts by law. Collective bargaining de…nes wage ‡oors for di¤erent categories of workers, but employers can deviate from this agreed wage level and pay higher wages. Restrictive employment legislation is a barrier to labour mobility, one of the focal points of this paper. Low levels of job mobility, in turn, create incentives for …rms to use …xed-term contracts and reduces the incentives for …rms to provide training.4 Nevertheless, in spite of the rigid legal setting and the nature of collective bargaining, Portugal was reported by OECD in 2003 to have one of the lowest unemployment rates in the EU and a high degree of wage ‡exibility in the private sector and when the country is faced with economic shocks (OECD, 2003). 4 Industry a¢liation and wages In the competitive model the wages of workers do not depend on …rm or industry a¢liation. This prediction is usually tested by de…ning a wage function as follows: yijt = x ;t + k(j(i;t)) + 3 4 ij Although merit promotions are totally dependent on the employer’s will. For more details see OECD (2006). 8 (1) where yijt is the logarithm of real monthly wages of worker i = 1; :::; N in …rm j = 1; :::; J in period t = 1; :::; T ; x ;t is the vector of observed time varying covariates of workers (i; t) and …rms (j; t); k = 1; :::; K is a vector of mutually exclusive dummy variables indicating the industry a¢liation of …rm j; and ij is the idiosyncratic error. and are the parameters to estimate. If wages do not depend on industry a¢liation, then the parameters should be jointly equal to zero. In what follows, we test this prediction. 4.1 Cross-sectional interindustry wage di¤erentials To assess the existence and stability of relative wages across industries we estimate model (1) using annual data for the period from 1986 to 2000. In all our speci…cations, we control for worker and …rm observed characteristics. Worker related variables include the type of job mobility experienced by the worker within the last year (automatic or merit promotion, entry to the …rm after a short or long period of non-employment), gender, years of seniority at the …rm and its square, years of potential labour market experience and its square, monthly hours of work and its square, education level (up to ISCED 0/1, ISCED 2, ISCED 3, ISCED 5/6), skill level split into 3 categories (low, medium, high), occupation (ISCO 9 major categories), and a dummy for part time or full time work. Firm-related covariates include percentage of foreign capital, size of …rm (micro, small, medium or large), legal structure of the …rm (public …rm - ruled by private sector laws, sole proprietor, anonymous partnership, limited liability company and other), instrument of collective regulation (4 categories), and region (20 categories). Macroeconomic conditions are controlled for by inclusion of year indicators. The central variable of our analysis is the industry a¢liation of the …rm. The classi…cation of industries in Portugal follows the European Standard Industrial Classi…cation (SIC codes). However, until 1994 (inclusive) the coding followed a revision approved in 1978, and from 1995 onwards the coding follows a revision approved in 1993. This change in the coding system makes harmonization over the period di¢cult and the highest level of disaggregation possible with the data results in 43 di¤erent industries. Descriptive statistics of the variables are presented in Table 1. [Table 1 about here] 9 The cross-sectional estimates of equation (1) for the period 1986-2000 are shown in Tables 2 and 3. These suggest the existence of important interindustry wage di¤erentials. Industry parameters are, in general, individually statistically signi…cant and, contrary to the competitive prediction, the hypothesis that the coe¢cients are simultaneously null is rejected (see F- statistics in the tables). Therefore, workers with the same observed characteristics working in …rms with observably equal characteristics have di¤erent wages depending on the industry in which they are employed. The wage di¤erences identi…ed are considerable, with some industries paying wages that are more than 41% above the economy average while others pay wages that are more than 26% below the average. Amongst the industries paying the lowest wages, we …nd the Manufacture of furniture, Textiles, Clothing and Restaurants and cafés, where the wage penalty ranges from 8% to 26% below the economy average. The industries paying the highest wages tend to be related to …nancial intermediation like Banking or Insurance, the Electricity, gas and water supply services, and productive services such as transport. The wage premia paid by these industries ranges between 8% to 41% above the economy average. There is considerable dispersion in this wage structure, the estimated standard deviations of the industry coe¢cients range from 8% to 11% in the period. This suggests that interindustry mobility can have a large impact on wages.5 [Tables 2 and 3 about here] The estimated industry wage structure, however, can be partially transitory and therefore not stable over time. That is, due to demand ‡uctuations some industries may lower wages relative to the wages paid elsewhere without leading workers to switch to expanding industries. We assess the role of transitory shocks by analysing the degree of linear association between the industry wage structure observed in one point in time, with that observed in other points in time. For this purpose we compute correlations between the 43 industry coe¢cients observed in each year with those observed in 1986. As is shown in Table 4, the industry wage structure was fairly stable in the period 1986-2000. Weighted correlations between relative wages in 1986 5 The magnitudes of the weighted standard deviations are comparable to those obtained by Krueger and Summers (1988) and Goux and Maurin (1999). 10 and all of the subsequent years range from 0.76 to 0.98 which suggests that the structure barely changed in the 15 year period analysed.6 Therefore, we conclude that temporary variations in productivity do not seem to drive the structure of wages between industries.7 [Table 4 about here] 4.2 Contribution of industry a¢liation to wage dispersion Taken with the previous results, we see that industry a¢liation is a signi…cant and stable determinant of wages. We now determine its relative importance in explaining observed wage dispersion using analysis of covariance. In model (1), the total proportion of wage variation explained by the covariates (X) and industry a¢liation (K) is given by the R2 of the regression. If X and K were not correlated, regressions of log wages on each of the covariates alone would give a unique decomposition of the contribution of each set of variables to the total explained variation. However, the possible collinearity between the two sets implies there is no unique variance decomposition. Nevertheless, we can identify the bounds of the share of variance explained by each set of variables. The share of wage variation unambiguously associated to K is given by the increase in the explanatory power arising from adding industry dummies to a wage regression already including X. This marginal contribution of K corresponds to the minimum estimate of the relative size of the variance contributed by K. The upper bound for the importance of the industry e¤ects is given by the R2 of a wage equation including only industry dummies. For this analysis (and that which follows) the observations for the period from 1986 to 2000 are pooled.8 The basic decomposition of the sources of wage dispersion is presented in Table 5. The proportion of the variance in wages explained by the covariates and industry a¢liation together is 72%. The covariates (industry) …rst speci…cation allows identi…cation of the portion of wage variation associated unambiguously to the industry (covariates) e¤ects. Therefore, the 6 There is a discrete fall (of about 0.05) in the strength of the correlation between 1994 and 1995, this fall is most likely to have been caused by the change in the industry coding system than by an e¤ective decrease in the correlations between the wage structure observed in the years post 1994 with that observed in 1986. 7 According to Krueger and Summers (1987), the structure persists and changes only moderately over longer intervals. 8 For all the speci…cations in which the data is pooled, the right hand side and left hand side variables are N P J P Tij y ij coded in deviations from the grand means, where the grand mean of yijt , e.g., is given by: y = : T i=1j=1 11 minimum estimates of the relative size of the variance contributed by industry and covariates is 2% and 43%, respectively. The upper bound is 29% for industry e¤ects and 70% for the covariates. The large range in the explanatory power (e.g. industry e¤ects account for between 2% and 29% of wage variation) arises from a large degree of collinearity between industry and the covariates. These results are in line with those obtained by Dickens and Katz (1987) for the U.S.A., who …nd that industry e¤ects account for between 7% to 30% of wage variation using 1983 CPS data. They suggest that industry a¢liation is an important factor explaining wage dispersion and that noncompetitive mechanisms may be at work in the labour market. [Table 5 about here] 4.3 Raw interindustry wage di¤erentials As mentioned previously, a non-competitive labour market is not the only possible explanation for the observed interindustry wage di¤erentials resulting from model (1). Our data set allows us to test whether the industry wage structure remains after we control for unobserved worker, …rm and match e¤ects. Therefore, we can investigate if true interindustry wage di¤erentials exist or if the di¤erentials observed in cross-sectional data simply re‡ect an unequal distribution of unmeasured heterogeneity across industries. Before considering the simultaneous impact of the three types of unobserved e¤ects, we …rst analyse, with the pooled data, how interindustry wage di¤erentials change as we gradually control for worker and …rm heterogeneity. These results are shown in Table 6, and each of its columns includes additional controls. Column (1) displays the results of a model that controls only for industry a¢liation and time e¤ects. This column gives us the unadjusted wage di¤erentials between industries, i.e., the di¤erence between average wages in the industry and the economy wide average of wages. Column (2) adds to the model observed worker characteristics, this column gives us interindustry wage di¤erentials adjusted for worker characteristics. In column (3) observed …rm characteristics are added as controls. The di¤erentials observed in this column are what we will, henceforth, call raw interindustry wage di¤erentials. That is, they summarize the industry wage structure adjusted for all the worker and …rm characteristics that we can observe in the data. Column (4) additionally controls for unobserved worker heterogeneity using the within-worker transformation. 12 From column (1) we conclude that, similar to our previous cross-sectional results, the industries paying the highest wages are Banking, and Insurance, Electricity, gas and water supply services for which the unadjusted wage premia are more than 66%. Those paying the lowest wages are Clothing, Furniture and Shoes manufactures, for which the unadjusted wage penalty is more than 40%. With no controls for any type of heterogeneity, the standard deviation of the estimated industry wage premia is 29%, which suggests that substantial wage growth can be realized by changing industries. However, the estimates are very sensitive to whether worker observed heterogeneity is controlled for. As we can see from column (2), the explanatory power of the model increases signi…cantly (from 33% to 68%) when we control for measured worker characteristics, while the dispersion of wages across industries is almost halved (the standard deviation is now 0.15). Adding …rm observed characteristics (column 3) further increases the explanatory power of the model (to 72%). While in some industries adding observed characteristics of …rms hardly a¤ects the estimated industry e¤ect (compared to the model in column 2), in some others it substantially reduces it. This is re‡ected in inter-industry wage dispersion, as the standard deviation of the interindustry wage e¤ects falls from 0.15 to 0.10. Thus, it seems that observed characteristics of workers and …rms explain much of the observed di¤erences in wages across industries.9 If unmeasured ability is time invariant and equally rewarded in all industries, then unobserved productive ability is an individual …xed e¤ect that disappears using the within-worker transformation.10 The estimates from the model that applies this transformation are presented in column (4). In this model the explanatory power of observed worker and …rm characteristics is greatly reduced: the R2 is now 47%. This suggests that unmeasured labour quality is correlated with measured characteristics and shows the importance of controlling for unobserved e¤ects. Furthermore, much less variation in wages across industries remains. The standard deviation of wages across industries is now 0.06, which means that workers who change industries experience small wage changes. These results are in line with those obtained, for example, by 9 Using the 2002 European Structure of Earnings Survey, Magda et al. (2008) …nd similar results for the dispersion of interindustry wage di¤erentials presented in columns 1, 2 and 3 of Table 6. In a group of 11 Eastern and Western European countries, Portugal was found to be one of the countries with highest dispersion in the interindustry wage structure when controls for observed worker and …rm characteristics are included. 10 The condition that ability is rewarded equally in every industry is what makes it a worker-speci…c …xed e¤ect, otherwise there is matching (Gibbons and Katz, 1992). If matching exists, the ability of the worker is of the same level in every industry, but the quality of the match (hence compensation) may di¤er across industries. 13 Goux and Maurin (1999) and Carruth et al. (2004) after controlling for person unobserved heterogeneity. [Table 6 about here] Our …ndings suggest that wages may be being set competitively but more able workers are concentrated in certain industries, making wages appear to be larger in some industries than in others. That is, industries paying higher wages are di¤erent from other industries in that they hire a higher proportion of high wage workers. This could drive us to conclude, as Murphy and Topel (1987), that most of the raw interindustry wage di¤erentials (in column 3) are due to unobserved worker heterogeneity and not to true di¤erences in …rm compensation policies across industries. However, so far, we have not accounted for unobserved characteristics of …rms, despite …rms being the wage determining units and industry being a characteristic of the …rm. If industry is a wage contour, then ignoring compensation policies of …rms in a study of interindustry wage di¤erentials is a major weakness.11 Moreover, as Goux and Maurin (1999, p. 506) suggest, "if the interindustry di¤erentials are not measured as an average of …rm e¤ects, an uncertainty over the correct interpretation of estimated industry e¤ects will persist." This means that the analysis cannot be complete until we disentangle the roles of workers and …rms in de…ning the industry wage structure. If …rms are an important factor in explaining interindustry wage di¤erences then, as Krueger and Summers (1987) suggest, the dispersion in wages between industries must be decomposed into three parts, the part due to person e¤ects, the part due to …rm e¤ects and the part due to the covariance between the two. To do this, we use AKM’s (1999) exact decomposition. 5 Statistical approach 5.1 Unobserved heterogeneity and interindustry wage di¤erentials In the previous section we speci…ed a competitive model of wage determination that only considers observed characteristics of workers and …rms, and unobserved worker characteristics 11 A wage contour is de…ned by Dunlop (1964, p. 17) as a stable group of …rms "[...] which are so linked together by (a) similarity of product markets, (b) by resort to similar sources of labour force, or (c) by common labour market organization that they have common wage making characteristics". 14 as determinants of wages. We now specify a match e¤ects model that considers not only observed characteristics, but also unmeasured worker, …rm and match e¤ects.12 This will allow the estimation and decomposition of interindustry wage di¤erences into parts attributable to unobserved individual, …rm, and worker-…rm match heterogeneity. The match e¤ects model estimates a wage equation of the type: y =X +D +F where X(N means); D(N Z) +G + (2) is the matrix of observable time varying covariates (in deviations from the grand N) is the matrix of indicators for worker i = 1; :::; N ; F(N J) indicators for the …rm at which worker i is employed at period t; and G(N of indicators of worker-…rm matches. y is a (N is the matrix of M) is the matrix 1) vector of log monthly real wages (also in deviations from the grand means).13 The set of parameters to estimate are ; the Z of coe¢cients on the covariates; , the N …rm e¤ects; and the M 1 vector of worker e¤ects; , the J 1 vector 1 vector of 1 vector of unobserved match e¤ects. Because industry is a characteristic of the …rm, the pure interindustry wage di¤erential, conditional on the same information as in equation (2), is de…ned as k for some industry classi…cation k = 1; :::; K.14 Therefore, the de…nition of the pure industry e¤ect ( k ) is the aggregation of the pure …rm e¤ects ( ) within the industry, that is k T N X X 1(K(J(i; t)) = k) Nk i=1 t=1 J(i;t) (3) where Nk J X 1(K(j) = k)Nj j=1 and K(j) is a function denoting the industry a¢liation of …rm j.15 This aggregation of J …rm 12 This section follows very closely the framework developed in section 2.1 in AKM (1999), and Woodcock’s (2008a) extension to incorporate match e¤ects. 13 In the presence of an unbalanced panel dataset (as we have here) where both workers and …rms can enter T P or exit the panel during the period of analysis, the total number of observations per worker is N = Ti: i=1 14 Model (1) does not consider …rm e¤ects. Therefore, it involves the aggregation of …rm e¤ects into industry dummy indicators. 15 These equations, e.g. correspond to those found in AKM (1999) p. 258. 15 e¤ects into (2), K(J(i;t)) , J(i;t) k industry e¤ects corresponds to including industry indicator variables in equation and de…ning what is left of the pure …rm e¤ect as a deviation from industry e¤ects, 16 K(J(i;t)) . In matrix notation: y = X + D + F A + (F where the matrix A, J (4) K, with element ajk = 1 if K(j) = k; classi…es each of the J …rms into one of the K industries. The parameter vector the pure …rm e¤ects. (F FA ) + G + , K 1, is the weighted average of F A ) is the …rm e¤ect net of industry e¤ects. This e¤ect can also be expressed as MF A F , where MF A is the matrix that obtains deviations from industry means. The least squares estimates of equation (4) have no biases due to omitted variables or to aggregation as (4) only decomposes F into two orthogonal components: the industry e¤ects F A , and the …rm e¤ects net of the industry e¤ect (F F A ). It is worth noticing that, because industry a¢liation is de…ned as a characteristic of the …rm, we do not have to actually run model (4). Pure industry e¤ects, F A , are standard averages of …rm e¤ects within the industry, as shown in (3). An alternative method of computing these averages (and to make the orthogonal decomposition of the pure …rm e¤ects) is to specify a model that regresses the pure …rm e¤ects (F ), estimated from equation (2), on the set of mutually exclusive dummy variables for the K industries.17 If wages are in fact determined according to speci…cation (2), that is if the expected values or probability limit of unobserved worker, …rm and match e¤ects are non-zero, then the estimated returns to the observed characteristics, industry a¢liation included, are biased if we use model (1).18 AKM (1999) discuss the biases that arise due to omitted residual …rm e¤ects (column 4 of Table 6), and to omitted person and residual …rm e¤ects (column 3 of Table 6) when 16 Authors attempt to use industry classi…cations as detailed as to have more than 90 industry codes. The reason for decomposing industrial aggregates into the most detailed level possible is related to the possibility that average compensation policies of …rms may vary across …ner levels of classi…cation and not within aggregates, and so estimates can be subject to aggregation biases. The pure industry e¤ects, however, are not subject to this bias because they are computed from …rm-level estimates. (Woodcock, 2008) 17 To clarify, we know that in a model without a constant: Yi = Xi + ui , Yi = E[Y jXi ] + ui . Therefore, the coe¢cients obtained are the pure industry e¤ects (F A ), or average …rm e¤ects within the industry, and the residual from this regression is the remaining, or residual, …rm e¤ect (F F A ): (This result is true because industry dummies are mutually exclusive and their covariance is zero.) 18 In the case where person, …rm and match e¤ects have non-zero expectation, the bias would not exist only if these components were orthogonal to the observed covariates, which is unlikely. 16 we specify the wage equation as a function of unobserved person and …rm heterogeneity only (that is, the match e¤ect (G ) is included in the error term). Woodcock (2008a) extends this discussion by deriving the biases caused by the omission of match e¤ects when wages are also determined by match unobserved heterogeneity. Understanding the nature and composition of these biases is the tool for decomposing the raw interindustry wage di¤erentials (column (3) of Table 6) into the contributions due to worker, …rm, and match e¤ects. Consequently, in the next section we revise the biases generated by the omission of the three components of unobserved heterogeneity, focussing solely on the industry coe¢cients, and explain the procedure to identify the relative importance of each unobserved e¤ect in explaining the raw interindustry wage di¤erentials.19 5.2 Omission of person, …rm and match e¤ects If the true data generation process is given by equation (2) but estimates of industry e¤ects are based upon a model that omits person, …rm and match e¤ects (and so we estimate some instead of the pure industry e¤ect, ), this implies that D ; (F , F A ) and G of model (4) are moved into the error term and our model takes the form y=X where " = (F + FA +" (5) F A ) + D + G + . Because the set of regressors can be broken up in two groups, in this case X (observed characteristics of workers and …rms) and F A (industry e¤ects), we can transform (5) as follows y = PW y + M W y = X where W = [X + FA + MW y (6) F A]; PW = X(X 0 X) 1 X 0 ; is the matrix that averages the observations across time for each individual and has typical element ui: ; and MW = I PW ; is the matrix that obtains the deviations from individual means and has typical element uit 19 The focus on industry coe¢cients only is for clarity of reasoning. Similarly, the the same biases and so this discussion also applies to them. 17 ui: . and are parameters su¤er from the least squares estimates obtained from (5). Premultiplying (6) by (F A)0 MX we obtain A 0 F 0 M X y = A0 F 0 M X F A and solving with respect to we obtain the estimator of industry e¤ects20 = (A0 F 0 MX F A) 1 A0 F 0 MX y: Under the assumption that (7) is uncorrelated with (but correlated with the other components of the error term of (5)), and because MX annihilates X (that is, MX X = 0); the expectation of (7) is E[ that is the estimator ]= + (A0 F 0 MX F A) 1 A0 F 0 MX (MF A F +D +G ) is equal to the pure industry e¤ects, , plus the sum of the employment- duration weighted average of the residual …rm e¤ect, the person and match e¤ects inside the industry, given X. This means that the bias is equal to the sum of the weighted portion of …rm, person and match e¤ects that is explained by the included covariates. Given that the pure …rm e¤ect, F , is equal to the sum of the pure industry e¤ect, F A ; with the residual …rm e¤ect, F F A ; we can rearrange the previous equation and obtain E[ ] = (A0 F 0 MX F A) 1 + (A0 F 0 MX F A) A0 F 0 M X F 1 + (A0 F 0 MX F A) 1 A0 F 0 M X D + (8) A0 F 0 M X G : This expression shows that the raw interindustry wage di¤erential, i.e. the di¤erential obtained in a model that does not include unobserved worker, …rm and match e¤ects, can be decomposed into the sum of the industry average …rm e¤ect, the industry average person e¤ect and the industry average match e¤ect, each of these averages conditional on X. These averages are the expectation of the least squares estimator in auxiliary regressions of each of the omitted 20 These estimates are obtained from a Frisch-Waugh-Lovell (FWL) regression. In a model in which the regressors can be split into two groups, and these are transformed to be mutually orthogonal, then OLS estimates of the parameter of interest obtained either from the original speci…cation or from the modi…ed model are numerically identical. See Davidson and MacKinnon (2004) for a thorough presentation of the FWL theorem. 18 regressors on the included regressors.21 Equation (8) is exact if the values of and and are known in which case we can have a consistent estimate of the decomposition based in (8). The decomposition is done in the following section. 6 The sources of interindustry wage di¤erentials In this section we decompose the raw interindustry wage di¤erentials into proportions due to person, …rm and match e¤ects, as shown in equation (8). The raw wage di¤erentials, are estimated using model (1). The person, ; …rm, ; ; and match, ; e¤ects are estimated from the match e¤ects speci…cation (2). The estimation of this model involves a three-step procedure. Firstly, is estimated after transforming (2) into deviations from match-speci…c means. Results from partitioned regression imply that b is a consistent estimate of . Secondly, b and b are computed using the person and …rm e¤ects model (a model without unobserved match e¤ects as in AKM, 1999).22 Finally, the match e¤ects estimator is de…ned as the error from the regression of worker-…rm matches on a constant and on the person and …rm e¤ects estimated previously.23 This match e¤ect can be correlated with the observed covariates, but is orthogonal to the person and …rm e¤ects. 6.1 Pure interindustry wage di¤erentials Before moving to the exact decomposition of the raw interindustry wage di¤erentials we present the resulting industry wage structure when we control only for person and …rm e¤ects, and when we also include match e¤ects in equation (2). We are thus able to establish whether true interindustry wage di¤erentials exist after we control for all types of measured and unmeasured heterogeneity. These results are shown in Table 7. To the extent that it adds further controls for unobserved heterogeneity, this table can be considered a continuation of Table 6. Compared 21 This is easier to recognise if we compare the expression of each component of equation (8), with the logic that connects equations (5) and (7). 22 These were estimated using the exact least squares solution (instead of the approximate method of AKM, 1999) as developed by Abowd, Creecy and Kramarz (2002), Ferreira (2009b). T ij P yijt xijt b 23 = (b u + bi + b j + bij ); where b is the withinThat is, we estimate the following model: b y ij = Tij t=1 match estimator of , and b and b are the worker and …rm unobserved e¤ects previously identi…ed using the person and …rm e¤ects model. 19 to a speci…cation that includes only unobserved worker characteristics (Table 6, column 4), the interindustry wage dispersion doubles when we also consider both person and …rm e¤ects (Table 7, column 1) or when we control for person, …rm and match e¤ects (Table 7, column 2). The results for the match e¤ects model are unbiased estimates of interindustry wage di¤erentials. The adjusted standard deviation is now 0.15, which suggests that unobserved abilities of workers are not the sole factor in‡uencing productivity, hence wages, and that compensation policies of …rms vary across industries. [Table 7 about here] The wage structure reported in Table 7 corresponds to our estimate of the pure interindustry wage di¤erentials. These results, however, are an average for the 1986-2000 period and do not allow us to verify the persistence of the pure interindustry wage structure over the 14 year period. To assess how stable these di¤erentials are, we computed the annual average of …rm e¤ects within industries and correlated the yearly industry coe¢cients with that observed in 1986. Our results reveal that the magnitude of the weighted correlations of the pure industry e¤ects in 1986 with those of the following years ranges from 0.92 in 2000 to 0.999 in 1987 (see Table 8). Therefore, the pure interindustry wage di¤erentials are unlikely to be caused by transitory shocks. Given the persistence and large dispersion of the pure interindustry wage structure we understand that if a worker changes industries he can experience considerable wage growth. [Table 8 about here] 6.2 Decomposition of raw interindustry wage di¤erentials We now proceed with the decomposition of the raw interindustry wage di¤erentials into a part due to person e¤ects, a part due to …rm e¤ects and another part due to match e¤ects as shown in equation (8). The raw industry e¤ects, , estimated from equation (1), are presented in column 3 of Table 6 for reference. The other components of equation (8), that is the industry average …rm e¤ect (the …rst component of the equation) the industry average person e¤ect (the second component), and the industry average match e¤ect (the third component) were 20 estimated using model (2). Results for the estimated version of equation (8), both when we are in the context of a person and …rm e¤ects model or in the context of a match e¤ects model, are presented in Table 9. These regressions yield an R2 of 1 and the coe¢cients on the industry average person and …rm e¤ects are very close to 1, see column (1). The same is not true for the coe¢cient of the average match e¤ects, suggesting less precision in its estimate.24 This means that these components fully account for the raw industry e¤ects and that the estimates of the interindustry wage di¤erentials are su¢ciently precise to allow an accurate decomposition. Since the industry average person, …rm, and match e¤ects are centered to have a zero sample mean, the terms high wage-worker, -…rm, or -match mean, respectively, workers, …rms or matches whose e¤ect is greater than the economy-wide average of zero. The same interpretation applies to the raw interindustry wage di¤erentials. The R2 in columns (2) through (4) show results for the upper bound of the share of person, …rm and match e¤ects, respectively, in explaining the raw interindustry wage di¤erentials. With the estimated version of equation (8) we …nd that industry average person e¤ects are able to explain at most 31% while the industry average …rm e¤ects explain a maximum of 90% of the observed dispersion in wages across industries.25 In Table 10 we present the exact decomposition of raw interindustry wage di¤erentials for both the person and …rm e¤ects model and the match e¤ects model. This decomposition is done in proportional terms as follows share of share of due to individual e¤ects = due to …rm e¤ects = (A0 F 0 MX F A) (A0 F 0 MX F A) 1 1 A0 F 0 M X D A0 F 0 M X F (a) (b) and share of due to match e¤ects = (A0 F 0 MX F A) 1 A0 F 0 MX G : (c) In the case of the person and …rm e¤ects model, (a) and (b) (shown in columns 2 and 3, respectively) must add to 1, whilst in the case of the match e¤ects model, (a), (b) and (c) 24 The estimated unobserved e¤ects should have an impact equal to its size in the regression, therefore, its coe¢cient should be one. 25 Note that, similar to AKM (1999), the person and …rm e¤ects are weakly correlated (correlation below 0.10) and so we expect little of the person e¤ect to be explained by the …rm e¤ect. 21 (shown in columns 4, 5 and 6, respectively) add up to 1. Regardless of the model used, the average proportion of raw interindustry wage di¤erentials due to industry average person e¤ects is about 30%, while the average proportion of raw di¤erentials due to the industry average …rm e¤ects is close to 70%. Match e¤ects have a smaller role in explaining the raw industry wage premia (3%). One could think that the low proportion explained by the match e¤ects is due to the assumption that they are orthogonal to person and …rm e¤ects. However, Woodcock (2008) also estimates negligible interindustry variation in the match e¤ects despite not assuming orthogonality. We have also checked whether the average proportion explained by each component would di¤er for the group industries paying wages above the economy average and those paying wages below the economy average. The results remained unchanged. Both in the case of high wage industries and low wage industries, industry average …rm e¤ects explain close to 70% of the estimated raw di¤erential. Our last exercise is to compute the correlation between the industry average person e¤ects and the industry average …rm e¤ects. Using the person and …rm e¤ects model we obtain a positive correlation (0.45) between these two components, which means that across industries we have either high wage workers working in high wage …rms, or low wage workers in low wage …rms. Therefore, from the person and …rm e¤ects model we conclude that the nature of the raw interindustry wage di¤erentials is related to positive assortative matching, and that the forces that sort person e¤ects are correlated with the forces that sort …rm e¤ects within industries. This result contrasts with that obtained by Abowd et al. (2005), who …nd weak positive correlations between industry average person and …rm e¤ects, and Woodcock (2008) who, when using the person and …rm e¤ects model, …nds this correlation to be weak and negative (-0.10). However, Woodcock (2008) …nds a positive correlation (0.60) between these two components when using estimates from the hybrid match e¤ects model. Our main …ndings can now be summarized as follows. We conclude that the raw interindustry wage di¤erentials are not a temporary disequilibrium in the labour market and are not due to systematic di¤erences in unobserved labour quality across industries. Di¤erences in compensation policies of …rms are the main source of the di¤erentials found in cross-sectional analysis. The pure interindustry wage structure (computed as the average …rm e¤ects within industry), on the other hand, shows considerable dispersion (weighted standard deviation of 22 0.15) and is also very persistent. To some extent, the results found seem to re‡ect the nature of the Portuguese labour market. Collective bargaining, the predominant means of wage negotiation in Portugal, on the one hand, helps generating wage contours at the industry level as most of the collective agreements are industry-wide and so cover …rms with di¤erent economic characteristics within the industry. This helps explaining the existence and persistence of a genuine wage structure across industries. On the other hand, we concluded that, in contrast to results found in other studies where both unobserved worker and …rm e¤ects were important in explaining the interindustry wage structure, the estimated wage di¤erentials in Portugal are mostly due to …rm e¤ects. This result is compatible with a labour market with low levels of labour mobility, but high levels of wage ‡exibility. As …rms can choose to pay higher wages to adjust for speci…c economic conditions, adjustments in the Portuguese labour market seem to be made via price rather than quantity, while in other economies adjustments may occur through mobility (worker e¤ect) and price (…rm e¤ect). 7 Testing the competitive model We have found that true interindustry wage di¤erentials exist in the Portuguese economy. Di¤erent …rms pay di¤erent wages to workers with the same characteristics (measured or unmeasured), and mobility across industries can generate substantial wage growth. However, it is not yet clear whether non-competitive forces are at work in the economy, or whether these di¤erentials are caused by mechanisms compatible with the competitive model such as rent-sharing or e¢ciency wage explanations. The hypothesis of rent sharing between …rms and workers is consistent with the fact that the wage premia received by workers in a particular industry extends over many occupations within the industry. It is also consistent with a positive relationship between pro…tability and wages (Blanch‡ower, Oswald and Sanfey 1996); and with a negative correlation between turnover and wage premia (Krueger and Summers, 1988). E¢ciency wage hypotheses are also consistent with a negative association between industry wage di¤erences and turnover. One mechanism that would generate this association is the existence of …rm (or industry) speci…c skills and 23 training. If some …rms have more speci…c skills than others and if they are providing training to their workers, then they are contributing to increase the productivity of its workforce, and making workers more costly to replace. This raises the threshold of the …rm in terms of labour turnover, and provides incentives to increase wages in order to reduce the likelihood of separation (Krause, 2000). If industry speci…c skills are more important than …rm speci…c skills, then all …rms within a certain industry will be acting in similar way. Hence, levelling up wages within the industry.26 If, on the other hand, the industry wage structure is generated by compensating di¤erentials for unobserved working conditions, no association is expected to be found between industry wage premia and quits. Therefore, the relationship between the industry wage premia and separations from …rms provides a test of the competitive model of industry wage determination. In this section we examine the association between the industry wage structure identi…ed in the previous section and the time workers take to separate from …rms. In the literature, this analysis is typically done using worker-initiated separations, that is quits. We cannot make such a distinction with our data. However, we have noticed in previous work (Ferreira, 2009b) that separations followed by short gaps (less than 1 year) of non-employment seem to be ruled by a process di¤erent from that ruling separations followed by long gaps (longer than 1 year) of non-employment. In particular, we concluded that short gaps is more likely to be formed by quits and long gaps by layo¤s. Hence, we use such a distinction in the present analysis of the association between interindustry wage di¤erentials and separations.27 We estimate parametric duration models where the dependent variable is the time to separate from …rms, the independent variable of interest is the pure industry e¤ect (the e¤ect obtained after controlling for all types of observed and unobserved heterogeneity), and the unit of observation is employment spells. 26 Parent (2000), e.g., …nds that industry speci…c skills are more important then …rm-speci…c skills in determining wage growth. 27 Although, for robustness checks we also estimate models considering all separations together and separations followed by long periods of non-employment. 24 7.1 Interindustry wage di¤erentials and labour mobility In this subsection we present the results obtained from estimating a duration model of the time to quit a …rm assuming that time follows a loglogistic distribution.28 As well as the industry wage premia estimated in the previous section, we also include the estimated residual …rm e¤ects, person and match e¤ects to control for worker unobserved ability, …rm compensation policies and match quality. The vector of observed covariates includes the age of the worker (and its square), gender, educational level, skill level, occupation, part-time or full-time work, dummy for having previously changed …rms, type of instrument of collective regulation, size of …rm, legal structure of the …rm, percentage of foreign capital, region and year. Estimates of the e¤ects of industry wage premia on the time to quit the …rm are presented in Table 11. Given the functional form of the model, these coe¢cients measure relative changes in survival time for a given absolute change in the regressors. A positive coe¢cient indicates that time to separation is lengthened, while a negative coe¢cient indicates that time to separation is shortened. Because industry e¤ects are measured in deviation from the grand means, these e¤ects are measured against an economy wide average of zero. The average e¤ect of interindustry wage di¤erentials on the time to separation is positive and signi…cant, which means that the larger the industry wage premium, the longer workers take to change …rms. In the case of quits, increasing the industry wage premium by one unit lengthens the time to quit the industry by 37%.29 The coe¢cient on the industry wage premium is also positive and signi…cant when we consider layo¤s (0.51) and all separations together (0.41).30 Our result is consistent with a context in which …rm or industry speci…c skills are important and where the gains of retaining workers for long periods outweigh the gains in productivity resulting from adjusting the labour force to every transitory demand shock. That is, …rms …nd it optimal to pay wages above the competitive level in order to provide incentives for workers 28 Where a quit is separation that is followed by a period of non-employment that lasted less than one year. The choice of loglogistic distribution follows from speci…cation tests. See Ferreira (2009a) for more details. 29 To clarify the interpretation, consider model (4) and that the estimated coe¢cient ( ) for industry k was zero. If the coe¢cient is now 1, i.e. if the industry pays wages 100% above the economy average, then the time to separate from a …rm in this industry, on average, increases by 37%. 30 We can think of two potential reasons for the larger e¤ect found for layo¤s than for quits. On the one hand, if a worker is …red, a signal about his quality might be being sent to other …rms within the industry, hence it takes longer to …nd a job. On the other hand, having worked for a high wage industry, workers have a high reservation wage. Hence, it takes longer to …nd a matching o¤er and/or to adjust expectations. 25 to remain at the …rm. If the bene…ts …rms extract from this longer term attachment o¤set the costs of higher wages, then this can be a pro…t maximising strategy. Therefore, the industry wage di¤erentials found in our data might not be totally incompatible with the competitive model. 8 Summary and conclusions This paper examines the sources of interindustry wage di¤erentials. In a competitive labour market, in the long run, homogeneous workers working in similar …rms are paid similar wages. Wage di¤erences across segments of the labour market, be it …rms or industries, are due to temporary di¤erences in productivity and are a signal for labour mobility. The ‡ow of workers across segments equalizes wages. However, cross-sectional analysis of wages typically …nd that there are signi…cant and persistent di¤erences in wages across industries. Unsurprisingly, our cross-sectional analysis using Portuguese data con…rms these results. The industry wage structure in Portugal shows high dispersion at points in time and the di¤erentials persist over time. These …ndings prompt the question of what explains the existence of a structure of wages across industries. The two most common explanations for the identi…ed cross-sectional interindustry wage di¤erentials are that the wage structure found is due to imperfect measures of labour quality; alternatively, genuine interindustry wage di¤erentials do exist and compensation policies of …rms vary across industries. If the industry wage premia are real, then noncompetitive mechanisms may be at work in the labour market, because by paying supracompetitive wages …rms might not be maximising their pro…ts. On the other hand, and compatible with a competitive model, …rms may …nd it pro…table to pay wages above the competitive level. We examined the sources of interindustry wage di¤erentials using panel data over a 15 year period. Our models control for observed worker and …rm characteristics, but also for unobserved worker, …rm and match heterogeneity. Results show that after controlling for all types of heterogeneity the true interindustry wage di¤erentials are sizeable and persistent. This means that compensation policies of …rms vary across industries and that by changing industries workers enjoy substantial wage growth. Furthermore, we conclude that …rm compensation 26 policies are the main source of the di¤erentials found in cross-sectional analysis and explain, on average, about 70% of the raw industry wage structure. Unmeasured worker abilities are not as important and account for less than a third of such structure. We thus conclude that interindustry wage di¤erentials are not a trick caused by unmeasured labour productive quality and are not a temporary disequilibrium in the labour market, they re‡ect di¤erent treats given by …rms across industries. Why might …rms treat their workers by paying them supra-competitive wages? We focus on the turnover strand of e¢ciency wage models to investigate if the industry wage structure is caused by mechanisms compatible with the competitive model, by testing the e¤ect of the pure wage premia on the time workers take to separate from …rms. Our results indicate that the e¤ect of the industry wage premium on the time to quit …rms is positive and signi…cant. This is consistent with a labour market in which industry speci…c skills are important and where e¢ciencies are gained from creating incentives to worker-…rm attachments. Some questions remain unanswered and demand further research into the black box of the …rm. The major driving force of these di¤erentials or why they exist, has yet to be established. Although e¢ciency wage models may provide reasons for …rms to pay higher wages, we still do not know what di¤erences in the production functions of …rms make labour more valuable in some industries than in others. If the reason is related to di¤erences in productivity, then we should be able to identify what generates productivity dispersion across industries. Also any association/causality between these wage contours and characteristics of the product market has yet to be determined. Answers to these questions enhance our understanding of the mechanisms determining why …rms pay nomcompetitive wages and why that happens with greater intensity in some industries than others. 27 References Abowd, John M., Robert H. Creecy and Francis Kramarz (2002) "Computing person and …rm e¤ects using linked longitudinal employer-employee data." U.S. Census Bureau Technical Paper No. TP-2002-06. Abowd, John M., Hampton Finer and Francis Kramarz (1999) "Individual and …rm heterogeneity in compensation: an analysis of matched longitudinal employer-employee data for the State of Washington." in Haltiwanger, J. C., J. I. Lane, J. R. Spletzer, J. J. M. Theeuwes and K. R. Troske (editors): The Creation and Analysis of Employer-Employee Matched Data. North-Holland, Amsterdam: 3-24. Abowd, John M., Francis Kramarz, Paul Lengermann and Sébastien Roux (2005) "Persistent interindustry wage di¤erences: Rent sharing and opportunity costs.", unpublished paper. Abowd, John M., Francis Kramarz and David N. Margolis (1999) "High wage workers and high wage …rms." Econometrica, 67(2):251-333. Blackburn, McKinley and David Neumark (1992) "Unobserved ability, e¢ciency wages , and interindustry wage di¤erentials." Quarterly Journal of Economics, 107(4): 1421-1436. Blanch‡ower, David G., Andrew J. Oswald and Peter Sanfey (1996) "Wages, pro…ts and rent-sharing." Quarterly Journal of Economics, 111(1): 227-251. Carruth, Alan, William Collier and Andy Dickerson (2004) "Interindustry wage di¤erences and individual heterogeneity." Oxford Bulletin of Economics and Statistics, 66(5): 811-846. Davidson, Russell and James G. MacKinnon (2004) Econometric Theory and Methods. Oxford University Press, New York. Dickens, William T. and Lawrence F. Katz (1987) "Interindustry wage di¤erences and industry characteristics." in Kevin Lang and Jonathan S. Leonard (editors): Unemployment and the Structure of Labour Markets. Basil Blackwell Inc., Oxford. 48-89. Dunlop, John T. (1964) "The task of contemporary wage theory." in John T. Dunlop (editor): The Theory of Wage Determination. Macmillan, London: 3-27. Ferreira, Priscila (2009a) "The determinants of promotions and …rm separations.", ISER Working paper No. 2009-11, University of Essex. Ferreira, Priscila (2009b) "Returns to job mobility: the role of observed an unobserved factors." ISER Working paper No. 2009-12, University of Essex. Gibbons, Robert and Lawrence Katz (1992) "Does unmeasured ability explain interindustry wage di¤erentials?" Review of Economic Studies, 59(3): 515-535. Goux, Dominique and Eric Maurin (1999) "Persistence of interindustry wage di¤erentials: A reexamination using matched worker-…rm panel data." Journal of Labor Economics, 17(3): 492-533. Krause, Michael U. (2002) "interindustry wage di¤erentials and job ‡ows." CentER working paper no. 2002-03. 28 Krueger, Alan B. and Lawrence H. Summers (1988) "E¢ciency wages and the interindustry wage structure." Econometrica, 56(2): 259-293. Krueger, Alan B. and Lawrence H. Summers (1987) "Re‡ections on the interindustry wage structure." in Kevin Lang and Jonathan S. Leonard (editors): Unemployment and the Structure of Labour Markets. Basil Blackwell Inc., Oxford: 17-47. Jovanovic, Boyan and Robert Mo¢t (1990) "An estimate of a sectoral model of labor mobility." Journal of Political Economy, 98(4): 827-852. Magda, Iga, François Rycx, Ilan Tojerow and Daphné Valsamis (2008) "Wage di¤erentials across sectors in Europe: an East-West comparison.", IZA DP No. 3830. Murphy, Kevin M. and Robert H. Topel (1987) "Unemployment, risk, and earnings." in Kevin Lang and Jonathan S. Leonard (editors): Unemployment and the Structure of Labour Markets. Basil Blackwell Inc., Oxford: 103-140. OECD [Organisation for Economic Co-operation and Development] (2006). OECD Economic Survey of Portugal, 2006. Paris: OECD. OECD [Organisation for Economic Co-operation and Development] (2003). OECD Economic Survey of Portugal, 2003. Paris: OECD. Parent, Daniel (2000) "Industry-speci…c capital and the wage pro…le: Evidence from the National Longitudinal Survey of Youth and the Panel Study of Income Dynamics." Journal of Labor Economics, 18(2): 306-323. Thaler, Richard H. (1989) "Anomalies: Interindustry wage di¤erentials." Journal of Economic Perspectives, 3(2): 181-193. Woodcock, Simon (2008) "Wage di¤erentials in the presence of unobserved worker, …rm and match heterogeneity." Labour Economics, 15(4): 772-794. Woodcock, Simon (2008a) "Match e¤ects", unpublished paper, Department of Economics, Simon Fraser University. 29 Tables Table 1: Descriptive statistics of variables Variable Log monthly real wage Seniority (years) Experience (years) Hours of work (monthly) Yearly gap Gender Men Women Education ISCED 1 ISCED 2 ISCED 3 ISCED 5/6 Occupations Directors Intellectual and scienti…c specialists Professional, technical (intermediate) Administrative and managerial workers Clerical and sales workers Agriculture, silviculture and …shing Production and related workers Equipment operators and labourers Unquali…ed workers Skill Level High Medium Low Type of work Full time Part time Type of job mobility Automatic promotion Merit promotion Separation, small gap Separation, big gap Size of …rm Micro Small Medium Large Instrument of collective regulation Collective agreement Collective contract Regulating law Firm agreement Legal structure of …rm Public (Private market law) Sole proprietor Anonymous partnership Limited liability company Mean 6.3 8.7 22 170 0.14 61.7 38.4 71.6 11.3 12.6 4.6 1.7 2.0 8.3 14.1 8.5 1.26 25.6 13.9 18.0 18.4 43.5 38.1 91.5 8.4 7.7 2.9 1.6 3.1 9.2 25.0 29.2 36.6 4.0 82.7 3.8 8.7 4.8 5.3 29.0 55.2 Variable Percentage of foreign capital Region Year Industry Agriculture Silviculture Fishing Mining Food products Beverages Tobacco Textiles Clothing Leather Shoes Wood and cork Furniture Pulp, paper, paper prod. Publishing and printing Industrial chemicals Other chemicals Petrol, rubber, plastics Ceramics Glass Other non-met min prod. Base metals Metallic products Non-electric materials Electric materials Motor vehicles Professional instruments Other manufacturing Elect., gas and water Building Wholesale trade Retail trade Restaurants and cafes Hotels Transport Communications Banking Insurance Real estate Productive services - transport Other productive services Social services Personal services Mean 9.1 20 Districts 1986-2000 1.20 0.09 0.21 0.73 3.92 0.67 0.09 7.45 6.53 0.33 3.05 2.01 1.45 0.88 1.33 0.64 1.01 1.42 1.30 0.52 1.60 0.92 3.38 1.80 2.19 2.09 0.25 0.39 1.31 9.38 7.28 7.84 2.83 1.87 5.05 1.97 3.06 0.92 3.73 0.84 0.79 3.66 2.00 Note: These statistics are computed over the sample of 1,823,572 worker-year observations. Source: Own calculations based on Quadros de Pessoal (1986-2000). 30 Table 2: Cross-sectional interindustry regression-adjusted wage di¤erences, 19861993 Industry Agriculture Silviculture Fishing Mining Food products Beverages Tobacco Textiles Clothing Leather Shoes Wood and cork Furniture Pulp, paper, paper prod. Publishing and printing Industrial chemicals Other chemicals Petrol, rubber, plastics Ceramics Glass Other non-met min prod. Base metals Metallic products Non-electric materials Electric materials Motor vehicles Professional instruments 1986 -0.186 (0.010) -0.071 (0.038) -0.135 (0.021) 0.031 (0.009) -0.064 (0.004) -0.095 (0.009) -0.102 (0.019) -0.081 (0.003) -0.124 (0.004) -0.005 (0.012) -0.087 (0.006) -0.120 (0.005) -0.239 (0.007) 0.047 (0.007) -0.041 (0.007) 0.073 (0.007) 0.043 (0.007) 0.009 (0.006) 0.021 (0.008) 0.205 (0.010) 0.003 (0.006) -0.013 (0.006) -0.066 (0.004) -0.073 (0.006) 0.056 (0.006) 0.008 (0.005) -0.031 (0.014) 1987 -0.20 (0.010) -0.046 (0.032) -0.114 (0.022) 0.007 (0.009) -0.073 (0.004) -0.104 (0.009) -0.077 (0.019) -0.075 (0.003) -0.095 (0.004) 0.021 (0.012) -0.071 (0.005) -0.109 (0.005) -0.246 (0.007) 0.053 (0.007) -0.046 (0.007) 0.067 (0.007) 0.034 (0.007) -0.012 (0.006) 0.037 (0.007) 0.205 (0.010) 0.013 (0.006) -0.026 (0.006) -0.047 (0.004) -0.070 (0.006) 0.065 (0.006) 0.043 (0.005) -0.026 (0.014) 1988 -0.167 (0.009) -0.041 (0.023) -0.181 (0.032) 0.025 (0.009) -0.078 (0.004) -0.107 (0.009) -0.028 (0.020) -0.108 (0.003) -0.108 (0.004) 0.018 (0.011) -0.084 (0.005) -0.114 (0.005) -0.247 (0.007) 0.036 (0.007) -0.049 (0.007) 0.105 (0.007) 0.081 (0.007) 0.091 (0.006) 0.041 (0.007) 0.168 (0.010) -0.000 (0.006) -0.010 (0.006) -0.055 (0.004) -0.051 (0.006) 0.084 (0.006) 0.005 (0.005) -0.061 (0.016) 1989 -0.153 (0.009) -0.054 (0.020) -0.271 (0.025) 0.033 (0.008) -0.089 (0.004) -0.120 (0.009) 0.046 (0.021) -0.115 (0.003) -0.112 (0.004) -0.008 (0.012) -0.108 (0.005) -0.095 (0.005) -0.229 (0.007) 0.038 (0.007) -0.031 (0.007) 0.094 (0.008) 0.058 (0.007) 0.115 (0.006) 0.029 (0.007) 0.152 (0.010) 0.035 (0.006) -0.012 (0.007) -0.047 (0.004) -0.048 (0.006) 0.067 (0.006) 0.005 (0.005) -0.048 (0.016) 1991 -0.106 (0.010) -0.012 (0.029) -0.185 (0.028) 0.076 (0.009) -0.076 (0.004) -0.123 (0.010) 0.117 (0.024) -0.119 (0.004) -0.113 (0.004) -0.028 (0.013) -0.135 (0.005) -0.083 (0.006) -0.234 (0.007) 0.075 (0.008) 0.019 (0.007) 0.052 (0.010) 0.058 (0.007) 0.101 (0.007) 0.033 (0.007) 0.118 (0.011) 0.054 (0.006) -0.023 (0.008) -0.039 (0.005) -0.033 (0.006) 0.073 (0.006) 0.087 (0.006) -0.044 (0.016) 1992 -0.126 (0.011) -0.075 (0.031) -0.217 (0.033) 0.099 (0.009) -0.078 (0.004) -0.112 (0.010) 0.003 (0.027) -0.123 (0.004) -0.128 (0.004) -0.006 (0.013) -0.114 (0.005) -0.096 (0.006) -0.219 (0.007) 0.034 (0.008) 0.014 (0.007) 0.066 (0.010) 0.067 (0.008) 0.073 (0.007) 0.029 (0.008) 0.098 (0.011) 0.041 (0.006) -0.016 (0.008) -0.028 (0.005) -0.030 (0.006) 0.074 (0.006) 0.078 (0.006) 0.015 (0.017) 1993 -0.122 (0.011) -0.023 (0.033) -0.243 (0.033) 0.112 (0.009) -0.069 (0.004) -0.106 (0.010) 0.034 (0.027) -0.138 (0.004) -0.149 (0.004) 0.002 (0.013) -0.114 (0.005) -0.091 (0.006) -0.261 (0.007) 0.030 (0.009) 0.033 (0.007) 0.086 (0.011) 0.058 (0.008) 0.074 (0.007) 0.011 (0.008) 0.099 (0.011) 0.025 (0.006) -0.004 (0.008) -0.019 (0.005) -0.010 (0.006) 0.071 (0.006) 0.086 (0.006) -0.049 (0.017) (Continued on next page) 31 Table 2: (continued from previous page) Industry Other manufacturing Elect., gas and water Building Wholesale trade Retail trade Restaurants and cafes Hotels Transport Communications Banking Insurance Real estate Productive services - transp Other productive services Social services Personal services F-stat Weighted SD No. of obs 1986 -0.095 (0.011) 0.258 (0.007) -0.052 (0.003) 0.014 (0.003) -0.049 (0.004) -0.091 (0.006) 0.031 (0.006) 0.109 (0.004) 0.091 (0.006) 0.289 (0.011) 0.413 (0.009) 0.102 (0.007) 0.287 (0.010) -0.096 (0.009) -0.088 (0.007) -0.088 (0.006) 227.81 0.104 103,925 1987 -0.098 (0.011) 0.246 (0.007) -0.068 (0.003) 0.009 (0.003) -0.045 (0.004) -0.110 (0.006) 0.020 (0.006) 0.126 (0.004) -0.021 (0.021) 0.266 (0.010) 0.398 (0.008) 0.092 (0.007) 0.348 (0.009) -0.108 (0.009) -0.094 (0.006) -0.068 (0.006) 250.50 0.103 104,893 1988 -0.112 (0.011) 0.232 (0.007) -0.074 (0.003) 0.012 (0.003) -0.048 (0.004) -0.127 (0.006) -0.001 (0.006) 0.110 (0.004) 0.131 (0.007) 0.305 (0.010) 0.360 (0.009) 0.064 (0.007) 0.267 (0.010) -0.130 (0.009) -0.086 (0.007) -0.068 (0.006) 231.79 0.104 109,632 1989 -0.100 (0.010) 0.247 (0.007) -0.060 (0.003) 0.012 (0.003) -0.037 (0.004) -0.121 (0.006) 0.020 (0.006) 0.108 (0.004) 0.022 (0.007) 0.369 (0.010) 0.348 (0.008) 0.041 (0.006) 0.335 (0.009) -0.149 (0.008) -0.111 (0.006) -0.055 (0.005) 275.61 0.108 119,886 1991 -0.084 (0.011) 0.245 (0.008) -0.068 (0.003) 0.020 (0.003) -0.034 (0.004) -0.121 (0.006) -0.006 (0.006) 0.050 (0.005) -0.016 (0.008) 0.243 (0.009) 0.263 (0.008) 0.017 (0.006) 0.277 (0.010) -0.146 (0.008) -0.100 (0.006) -0.049 (0.005) 216.49 0.102 128,766 1992 -0.054 (0.012) 0.259 (0.008) -0.076 (0.003) 0.030 (0.003) -0.028 (0.004) -0.120 (0.006) -0.028 (0.006) 0.045 (0.005) 0.019 (0.007) 0.295 (0.010) 0.313 (0.008) 0.041 (0.006) 0.254 (0.010) -0.143 (0.008) -0.093 (0.006) -0.038 (0.006) 224.71 0.109 132,284 1993 -0.057 (0.013) 0.236 (0.008) -0.086 (0.003) 0.029 (0.003) -0.009 (0.004) -0.130 (0.006) 0.009 (0.007) 0.047 (0.005) 0.046 (0.008) 0.265 (0.010) 0.255 (0.008) 0.030 (0.006) 0.259 (0.010) -0.127 (0.007) -0.065 (0.006) -0.029 (0.006) 220.92 0.107 130,095 Note: Because the model does not include a constant the resulting coe¢cients are proportionate di¤erences in wages between a worker in a given industry and the average worker in the economy. Standard errors in parentheses. Weights are industry employment shares for each year. Source: Own calculations based on Quadros de Pessoal (1986-2000). 32 Table 3: Cross-sectional interindustry regression-adjusted wage di¤erences, 19942000 Industry Agriculture Silviculture Fishing Mining Food products Beverages Tobacco Textiles Clothing Leather Shoes Wood and cork Furniture Pulp, paper, paper prod. Publishing and printing Industrial chemicals Other chemicals Petrol, rubber, plastics Ceramics Glass Other non-met min prod. Base metals Metallic products Non-electric materials Electric materials Motor vehicles Professional instruments 1994 -0.140 (0.011) -0.011 (0.031) 0.121 (0.031) 0.046 (0.010) -0.086 (0.004) -0.104 (0.010) 0.056 (0.029) -0.158 (0.004) -0.181 (0.004) 0.036 (0.013) -0.133 (0.005) -0.109 (0.006) -0.248 (0.007) 0.049 (0.009) 0.007 (0.007) 0.091 (0.013) 0.074 (0.009) 0.059 (0.007) -0.030 (0.008) 0.097 (0.011) 0.038 (0.006) 0.015 (0.009) -0.033 (0.005) -0.021 (0.007) 0.099 (0.006) 0.067 (0.007) -0.122 (0.017) 1995 -0.110 (0.011) -0.083 (0.026) -0.051 (0.022) 0.060 (0.009) -0.086 (0.004) -0.072 (0.009) 0.059 (0.031) -0.162 (0.004) -0.174 (0.004) 0.046 (0.014) -0.131 (0.005) -0.077 (0.006) -0.221 (0.006) 0.062 (0.009) -0.008 (0.007) 0.120 (0.012) 0.083 (0.008) 0.052 (0.007) -0.006 (0.007) 0.069 (0.011) 0.028 (0.006) -0.018 (0.010) -0.054 (0.005) -0.004 (0.006) 0.048 (0.006) 0.078 (0.006) -0.074 (0.015) 1996 -0.129 (0.011) -0.024 (0.025) -0.089 (0.023) 0.077 (0.010) -0.085 (0.004) -0.104 (0.010) 0.047 (0.032) -0.165 (0.004) -0.173 (0.004) 0.084 (0.013) -0.138 (0.005) -0.074 (0.006) -0.198 (0.007) 0.063 (0.009) -0.009 (0.007) 0.105 (0.011) 0.088 (0.008) 0.040 (0.007) -0.040 (0.007) 0.111 (0.011) 0.039 (0.007) -0.008 (0.010) -0.047 (0.005) 0.008 (0.006) 0.010 (0.005) 0.047 (0.006) -0.060 (0.014) 1997 -0.118 (0.010) -0.013 (0.023) -0.095 (0.022) 0.066 (0.009) -0.090 (0.004) -0.079 (0.010) 0.034 (0.031) -0.163 (0.004) -0.169 (0.004) 0.057 (0.014) -0.116 (0.005) -0.066 (0.006) -0.193 (0.006) 0.098 (0.009) 0.012 (0.007) 0.095 (0.011) 0.076 (0.008) 0.012 (0.007) -0.017 (0.007) 0.084 (0.011) 0.033 (0.006) 0.004 (0.010) -0.034 (0.005) 0.017 (0.006) 0.018 (0.005) 0.048 (0.006) -0.079 (0.015) 1998 -0.109 (0.10) -0.019 (0.023) -0.150 (0.022) 0.082 (0.009) -0.090 (0.004) -0.051 (0.010) 0.104 (0.029) -0.158 (0.004) -0.160 (0.004) 0.100 (0.015) -0.109 (0.005) -0.058 (0.006) -0.169 (0.006) 0.084 (0.008) 0.001 (0.006) 0.060 (0.011) 0.065 (0.008) 0.037 (0.007) -0.025 (0.006) 0.138 (0.010) 0.025 (0.006) -0.006 (0.009) -0.024 (0.004) 0.021 (0.005) 0.050 (0.005) 0.044 (0.005) -0.106 (0.014) 1999 -0.090 (0.010) 0.055 (0.022) -0.071 (0.023) 0.084 (0.009) -0.097 (0.004) -0.055 (0.010) 0.181 (0.027) -0.178 (0.004) -0.174 (0.004) 0.020 (0.015) -0.129 (0.005) -0.047 (0.006) -0.180 (0.006) 0.078 (0.009) 0.010 (0.006) 0.062 (0.011) 0.059 (0.008) 0.026 (0.007) -0.021 (0.006) 0.158 (0.010) 0.037 (0.006) -0.017 (0.009) -0.020 (0.004) 0.027 (0.006) 0.031 (0.005) 0.005 (0.005) -0.040 (0.014) 2000 -0.078 (0.010) 0.076 (0.024) -0.172 (0.023) 0.102 (0.009) -0.103 (0.004) -0.042 (0.010) 0.048 (0.027) -0.151 (0.004) -0.156 (0.004) 0.068 (0.016) -0.123 (0.005) -0.026 (0.006) -0.165 (0.006) 0.043 (0.009) 0.026 (0.006) 0.073 (0.012) 0.064 (0.008) 0.047 (0.007) -0.090 (0.006) 0.140 (0.011) 0.045 (0.006) 0.022 (0.010) 0.004 (0.004) 0.064 (0.006) 0.008 (0.006) 0.023 (0.005) -0.073 (0.014) (Continued on next page) 33 Table 3: (continued from previous page) Industry Other manufacturing Elect., gas and water Building Wholesale trade Retail trade Restaurants and cafes Hotels Transport Communications Banking Insurance Real estate Productive services - transp Other productive services Social services Personal services F-Stat Weighted SD No. of obs 1994 -0.071 (0.013) 0.301 (0.009) -0.086 (0.003) 0.031 (0.004) -0.029 (0.004) -0.151 (0.006) 0.001 (0.006) 0.021 (0.005) 0.071 (0.008) 0.235 (0.010) 0.223 (0.009) 0.016 (0.005) 0.217 (0.011) -0.160 (0.007) -0.069 (0.006) -0.032 (0.005) 227.19 0.110 130,439 1995 -0.107 (0.012) 0.277 (0.008) -0.088 (0.003) 0.009 (0.003) -0.025 (0.003) -0.168 (0.005) -0.013 (0.006) -0.018 (0.005) 0.131 (0.007) 0.238 (0.009) 0.279 (0.009) -0.072 (0.004) 0.214 (0.008) 0.047 (0.014) -0.060 (0.006) -0.018 (0.007) 238.22 0.108 134,981 1996 -0.103 (0.013) 0.230 (0.008) -0.089 (0.003) 0.009 (0.003) -0.024 (0.003) -0.181 (0.005) -0.012 (0.006) -0.016 (0.005) 0.128 (0.007) 0.251 (0.009) 0.285 (0.008) -0.075 (0.004) 0.238 (0.008) 0.062 (0.013) -0.058 (0.006) -0.022 (0.007) 241.78 0.108 134,284 1997 -0.111 (0.013) 0.239 (0.008) -0.084 (0.003) 0.003 (0.003) -0.017 (0.003) -0.166 (0.005) -0.027 (0.006) 0.020 (0.004) 0.130 (0.006) 0.191 (0.009) 0.300 (0.009) -0.074 (0.004) 0.205 (0.008) 0.047 (0.012) -0.068 (0.005) -0.010 (0.006) 233.22 0.099 142,809 1998 -0.113 (0.013) 0.200 (0.008) -0.088 (0.003) 0.002 (0.003) -0.026 (0.003) -0.154 (0.005) -0.038 (0.005) 0.023 (0.004) 0.087 (0.006) 0.157 (0.009) 0.282 (0.008) -0.069 (0.004) 0.167 (0.007) 0.037 (0.011) -0.058 (0.005) 0.013 (0.005) 223.73 0.090 146,159 1999 -0.130 (0.013) 0.199 (0.008) -0.081 (0.003) -0.002 (0.003) -0.024 (0.003) -0.157 (0.005) -0.034 (0.006) 0.050 (0.004) 0.052 (0.006) 0.165 (0.008) 0.257 (0.008) -0.079 (0.004) 0.104 (0.007) 0.026 (0.011) -0.067 (0.005) 0.006 (0.006) 224.50 0.091 150,921 2000 -0.081 (0.013) 0.103 (0.009) -0.061 (0.003) 0.026 (0.003) -0.007 (0.003) -0.135 (0.005) -0.015 (0.005) 0.070 (0.004) 0.037 (0.006) 0.117 (0.008) 0.281 (0.008) -0.057 (0.003) 0.083 (0.007) 0.016 (0.010) -0.036 (0.004) -0.015 (0.005) 200.48 0.08 154,498 Note: Given that the model does not include a constant the resulting coe¢cients are proportionate di¤erences in wages between a worker in a given industry and the average worker in the economy. Standard errors in parentheses. Weights are industry employment shares for each year. Source: Own calculations based on Quadros de Pessoal (1986-2000). 34 Table 4: Persistence of the raw interindustry wage structure (k ) between 1986 and 2000 Year 1986 1987 1988 1989 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 Note: Weighted correlation with 1986 1.00 0.979 0.979 0.962 0.921 0.938 0.914 0.891 0.839 0.839 0.838 0.822 0.807 0.761 Weights are average employment shares of each industry in the period from 1986 to 2000. Source: Own calculations based on Quadros de Pessoal (1986-2000). Table 5: Analysis of sources of wage variation Source of wage variation Model with covariates and industry (A) Error (1-A) Model with covariates …rst: Covariates (B) Industry (A-B) Model with industry …rst: Industry (C) Covariates (A-C) Variance of log wages Mean of log wages Total no. of observations No. of industries No. of covariates Share of TSS 0.718 0.282 0.699 0.019 0.292 0.426 0.276 0 1,823,572 43 74 Note: TSS stands for total sum of squares. Source: Own calculations based on Quadros de Pessoal (1986-2000). 35 Table 6: Estimated interindustry wage di¤erentials with di¤ering controls, pooled data Industry Agriculture Silviculture Fishing Mining Food products Beverages Tobacco Textiles Clothing Leather Shoes Wood and cork Furniture Pulp, paper, paper prod. Publishing and printing Industrial chemicals Other chemicals Petrol, rubber, plastics Ceramics Glass Other non-met min prod. Base metals Metallic products Non-electric materials Electric materials Motor vehicles Industry Add worker & time obs. e¤ects (1) -0.456 (0.003) -0.348 (0.011) -0.220 (0.007) 0.007 (0.004) -0.180 (0.002) 0.042 (0.004) 0.360 (0.010) -0.283 (0.001) -0.444 (0.001) -0.193 (0.005) -0.403 (0.002) -0.265 (0.002) -0.430 (0.003) 0.177 (0.003) 0.046 (0.003) 0.400 (0.004) 0.276 (0.003) 0.128 (0.003) -0.170 (0.003) 0.179 (0.004) -0.020 (0.003) 0.100 (0.003) -0.138 (0.002) -0.007 (0.002) 0.105 (0.002) 0.169 (0.002) (2) -0.234 (0.003) -0.117 (0.007) -0.075 (0.007) 0.030 (0.003) -0.105 (0.001) -0.060 (0.003) 0.175 (0.007) -0.181 (0.001) -0.193 (0.001) -0.049 (0.004) -0.148 (0.001) -0.153 (0.002) -0.306 (0.002) 0.089 (0.002) -0.038 (0.002) 0.162 (0.003) 0.092 (0.002) 0.053 (0.002) -0.023 (0.002) 0.155 (0.003) -0.003 (0.002) 0.008 (0.002) -0.088 (0.001) -0.038 (0.002) 0.123 (0.002) 0.081 (0.002) Add …rm obs. e¤ects, (3) -0.144 (0.003) -0.030 (0.007) -0.112 (0.007) 0.070 (0.002) -0.082 (0.001) -0.093 (0.003) 0.021 (0.007) -0.126 (0.001) -0.140 (0.001) 0.028 (0.004) -0.112 (0.001) -0.081 (0.002) -0.213 (0.002) 0.056 (0.002) -0.000 (0.002) 0.082 (0.003) 0.064 (0.002) 0.051 (0.002) -0.004 (0.002) 0.135 (0.003) 0.032 (0.002) -0.010 (0.002) -0.034 (0.001) -0.010 (0.002) 0.057 (0.002) 0.046 (0.002) Add worker unobs. e¤ects (4) -0.091 (0.004) -0.035 (0.008) -0.003 (0.008) 0.053 (0.004) -0.040 (0.002) -0.025 (0.004) 0.063 (0.019) -0.045 (0.002) -0.073 (0.002) -0.017 (0.006) -0.103 (0.003) -0.063 (0.003) -0.111 (0.003) 0.000 (0.004) -0.025 (0.003) 0.061 (0.004) 0.004 (0.003) 0.035 (0.003) 0.023 (0.004) 0.107 (0.006) 0.034 (0.003) -0.026 (0.003) -0.042 (0.002) -0.016 (0.002) 0.033 (0.002) 0.002 (0.002) (Continued on next page) 36 Table 6: (continued from previous page) Industry Professional instruments Other manufacturing Elect., gas and water Building Wholesale trade Retail trade Restaurants and cafes Hotels Transport Communications Banking Insurance Real estate Productive services - transp Other productive services Social services Personal services R2 Weighted SD Industry Add worker & time obs. e¤ects (1) -0.009 (0.006) -0.249 (0.005) 0.656 (0.003) -0.184 (0.001) 0.056 (0.001) -0.163 (0.001) -0.418 (0.002) -0.063 (0.002) 0.278 (0.002) 0.509 (0.002) 0.719 (0.002) 0.663 (0.003) -0.095 (0.002) 0.356 (0.003) -0.339 (0.004) -0.127 (0.002) -0.024 (0.002) 0.33 0.287 (2) -0.017 (0.004) -0.147 (0.003) 0.405 (0.002) -0.120 (0.001) -0.020 (0.001) -0.087 (0.001) -0.213 (0.001) 0.006 (0.002) 0.147 (0.001) 0.263 (0.002) 0.384 (0.001) 0.340 (0.002) -0.026 (0.001) 0.204 (0.002) -0.085 (0.002) -0.134 (0.001) -0.055 (0.002) 0.68 0.153 Add …rm obs. e¤ects, (3) -0.050 (0.004) -0.093 (0.003) 0.251 (0.002) -0.074 (0.001) 0.016 (0.001) -0.024 (0.001) -0.141 (0.001) -0.008 (0.002) 0.056 (0.001) 0.083 (0.002) 0.250 (0.002) 0.309 (0.002) -0.042 (0.001) 0.205 (0.002) -0.087 (0.002) -0.077 (0.001) -0.025 (0.002) 0.72 0.100 Add worker unobs. e¤ects (4) -0.011 (0.007) -0.046 (0.004) 0.117 (0.007) -0.050 (0.001) -0.007 (0.001) -0.035 (0.001) -0.102 (0.002) -0.003 (0.003) 0.041 (0.002) 0.086 (0.005) 0.184 (0.004) 0.236 (0.007) -0.032 (0.002) 0.051 (0.003) -0.021 (0.003) -0.051 (0.003) -0.056 (0.002) 0.47 0.064 Note: Because the model does not include a constant the resulting coe¢cients are proportionate di¤erences in wages between a worker in a given industry and the average worker in the economy. Standard errors in parentheses. Weights are industry average shares of employment in the period 1986-2000. The no. of observations is 1,823,572. Source: Own calculations based on Quadros de Pessoal (1986-2000). 37 Table 7: Estimated interindustry wage di¤erentials with di¤ering controls Industry e¤ect given X and: person and …rm e¤ects, Industry Agriculture match e¤ects, (1) -0.244 (0.001) -0.167 (0.005) -0.010 (0.004) 0.025 (0.002) -0.090 (0.001) -0.027 (0.002) 0.202 (0.005) -0.123 (0.001) -0.186 (0.001) -0.024 (0.003) -0.216 (0.001) -0.135 (0.001) -0.291 (0.001) 0.100 (0.002) -0.044 (0.001) 0.203 (0.002) 0.116 (0.002) 0.101 (0.001) -0.056 (0.001) 0.186 (0.002) 0.004 (0.001) 0.037 (0.002) -0.114 (0.001) -0.081 (0.001) 0.090 (0.001) 0.053 (0.001) Silviculture Fishing Mining Food products Beverages Tobacco Textiles Clothing Leather Shoes Wood and cork Furniture Pulp, paper, paper prod. Publishing and printing Industrial chemicals Other chemicals Petrol, rubber, plastics Ceramics Glass Other non-met min prod. Base metals Metallic products Non-electric materials Electric materials Motor vehicles person, …rm and (2) -0.253 (0.002) -0.173 (0.006) -0.010 (0.004) 0.025 (0.002) -0.093 (0.001) -0.028 (0.002) 0.209 (0.006) -0.127 (0.001) -0.193 (0.001) -0.025 (0.003) -0.223 (0.001) -0.140 (0.001) -0.301 (0.001) 0.104 (0.002) -0.046 (0.001) 0.210 (0.002) 0.120 (0.002) 0.104 (0.001) -0.058 (0.001) 0.193 (0.002) 0.004 (0.001) 0.039 (0.002) -0.118 (0.001) -0.084 (0.001) 0.093 (0.001) 0.055 (0.001) (Continued on next page) 38 Table 7: (continued from previous page) Industry e¤ect given X and: person and …rm e¤ects, Industry Professional instruments (1) -0.044 (0.003) -0.107 (0.003) 0.316 (0.001) -0.145 (0.001) -0.017 (0.001) -0.106 (0.001) -0.223 (0.001) -0.028 (0.001) 0.101 (0.001) 0.212 (0.001) 0.383 (0.001) 0.357 (0.002) -0.035 (0.001) 0.201 (0.002) -0.055 (0.002) -0.072 (0.001) -0.050 (0.001) -0.039 (0.000) 0.92 0.145 Other manufacturing Elect., gas and water Building Wholesale trade Retail trade Restaurants and cafes Hotels Transport Communications Banking Insurance Real estate Productive services - transp Other productive services Social services Personal services Constant R2 Weighted SD person, …rm and match e¤ects, (2) -0.045 (0.003) -0.111 (0.003) 0.328 (0.001) -0.150 (0.001) -0.018 (0.001) -0.110 (0.001) -0.231 (0.001) -0.029 (0.001) 0.105 (0.001) 0.220 (0.001) 0.397 (0.001) 0.370 (0.002) -0.036 (0.001) 0.208 (0.002) -0.057 (0.002) -0.075 (0.001) -0.052 (0.001) -0.041 (0.000) 0.94 0.150 Note: Because the model does not include a constant the resulting coe¢cients are proportionate di¤erences in wages between a worker in a given industry and the average worker in the economy. Standard errors in parentheses. Weights are industry average shares of employment in the period 1986-2000. The no. of observations is 1,823,572. Source: Own calculations based on Quadros de Pessoal (1986-2000). 39 Table 8: Persistence of the pure interindustry wage structure (k) between 1986 and 2000 Year 1986 1987 1988 1989 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 Note: Weighted correlation with 1986 1.00 0.999 0.998 0.995 0.989 0.988 0.984 0.978 0.945 0.943 0.935 0.935 0.923 0.922 Weights are average employment shares of each industry in the period from 1986 to 2000. Source: Own calculations based on Quadros de Pessoal (1986-2000). Table 9: Estimates of the relation between the raw interindustry wage structure and industry average person, …rm and match e¤ects Independent variables Person and …rm e¤ects model: Industry average person e¤ect Industry average …rm e¤ect R2 Match e¤ects model: Industry average person e¤ect Industry average …rm e¤ect Industry average match e¤ect R2 (1) 0.996 (0.007) 0.998 (0.002) 1.000 1.057 (0.012) 1.047 (0.009) 2.343 (0.257) 1.000 Coe¢cients: (2) (3) 1.692 (0.391) 1.100 (0.058) 0.313 0.899 (4) 1.618 (0.374) 1.061 (0.055) 0.313 0.899 -25.014 (0.591) 0.978 Note: These are coe¢cients of a regression of raw interindustry wage di¤erential on person and …rm e¤ects (for the person and …rm e¤ects model), and person, …rm and match e¤ects (for the match e¤ects model). Standard errors in parentheses. Source: Own calculations based on Quadros de Pessoal (1986-2000). 40 Table 10: Sources of raw inter industry wage di¤erentials, exact decomposition Pooled OLS Raw di¤erential ( Industry Agriculture Silviculture Fishing Mining Food products Beverages Tobacco Textiles Clothing Leather Shoes Wood and cork Furniture Pulp, paper, paper prod. Publishing and printing Industrial chemicals Other chemicals Petrol, Rubber, Plastics Ceramics Glass Other non-met min prod. Base metals Metallic products Non-electric materials Electric materials Motor vehicles Professional instruments Other manufacturing Elect., gas and water Building Wholesale trade Retail trade Restaurants and cafes Hotels Transport Communications Banking Insurance Real estate Productive serv - transp. Other productive serv. Social services Personal services Average Proportion ) (1) -0.144 -0.030 -0.112 0.070 -0.082 -0.093 0.021 -0.126 -0.140 0.028 -0.112 -0.081 -0.213 0.056 -0.000 0.082 0.064 0.051 -0.004 0.135 0.032 -0.010 -0.034 -0.010 0.057 0.046 -0.050 -0.093 0.251 -0.074 0.016 -0.024 -0.141 -0.008 0.056 0.083 0.250 0.309 -0.042 0.205 -0.087 -0.077 -0.025 Person and …rm e¤ects model, proportion of Person e¤. (2) 0.007 0.348 0.709 0.039 0.361 0.209 0.362 0.465 0.309 0.344 0.164 0.292 0.135 0.135 0.500 0.222 0.120 0.322 0.329 0.223 0.227 0.561 0.231 0.413 0.244 0.398 0.186 0.526 0.423 0.075 0.789 0.283 0.105 0.367 0.616 0.584 0.263 0.188 0.143 0.220 0.184 0.132 0.328 30.4% due to: Firm e¤. (3) 0.993 0.652 0.291 0.961 0.639 0.791 0.638 0.535 0.691 0.656 0.836 0.708 0.865 0.865 0.500 0.778 0.880 0.678 0.671 0.777 0.773 0.439 0.769 0.587 0.756 0.602 0.814 0.474 0.577 0.925 0.211 0.717 0.895 0.633 0.384 0.416 0.737 0.812 0.857 0.780 0.816 0.868 0.672 69.6% Match e¤ects model proportion of due to: Person e¤. Firm e¤. Match e¤. (4) 0.007 0.347 0.680 0.0379 0.349 0.203 0.362 0.450 0.300 0.342 0.162 0.284 0.131 0.133 0.499 0.220 0.117 0.321 0.329 0.221 0.225 0.562 0.230 0.414 0.237 0.387 0.184 0.510 0.409 0.073 0.756 0.281 0.102 0.366 0.600 0.563 0.255 0.182 0.142 0.212 0.173 0.130 0.329 29.8% (5) 0.960 0.644 0.276 0.929 0.613 0.762 0.633 0.513 0.664 0.648 0.817 0.682 0.834 0.843 0.494 0.763 0.852 0.670 0.666 0.761 0.759 0.435 0.759 0.582 0.727 0.580 0.797 0.455 0.554 0.900 0.200 0.706 0.863 0.626 0.368 0.398 0.709 0.782 0.843 0.748 0.762 0.845 0.666 67.7% (6) 0.033 0.009 0.044 0.033 0.037 0.035 0.005 0.037 0.036 0.010 0.021 0.034 0.035 0.024 0.007 0.016 0.030 0.009 0.005 0.018 0.016 0.003 0.011 0.004 0.036 0.033 0.019 0.035 0.038 0.027 0.044 0.014 0.035 0.008 0.036 0.039 0.036 0.036 0.014 0.040 0.065 0.025 0.005 2.5% Note: column (1) is a transcription of column (3) of Table 6, it reports the estimated raw interindustry wage di¤erentials for reference. Columns (2)-(6) report the proportional decomposition as described in subsection 6.2. Source: Own calculations based on Quadros de Pessoal (1986-2000). 41 Table 11: Interindustry wage di¤erentials and time to separation Industry premia No. of obs % separating Quits (small gap) 0.366 (0.063) Layo¤s (big gap) 0.512 (0.050) All separations 0.410 (0.040) 639,829 6.20 639,533 11.84 639,829 18.03 Note: Other covariates were included in the speci…cation, but their coe¢cients are not reported here. The average industry e¤ect is estimated from a regression in which the pure interindustry wage di¤erential is included as a continuous variable. Source: Own calculations based on Quadros de Pessoal (1986-2000). 42