Rigid wages and flexible labour? Firm-level evidence based on productivity for Belgium

Transcription

1 Rigid wages and flexible labour? Firm-level evidence based on productivity for Belgium Catherine Fuss and Ladislav Wintr June 6, 2008 Abstract: Based on firm-level data for Belgium over the period , we evaluate the elasticity of labour and average real wage per firm to microeconomic total factor productivity (TFP). We find that the elasticity of wages with respect to TFP is rather low. Our results suggest that sector-level collective bargaining in Belgium leaves limited margin for adjustment of real wages to firm-specific shocks. Next, we report evidence of a positive relationship between hours and technology. Finally, our finding of sluggish wage response in the services sector can be related to the observed higher price stickiness for services than for other goods. Keywords: wages, employment, hours, productivity. technology JEL: J30, J60 National Bank of Belgium corresponding author: Catherine Fuss, National Bank of Belgium, Research Department, 14, bd. de Berlaimont, 1000 Brussels, catherine.fuss@nbb.be The opinions expressed in this paper are solely our own and do not necessarily reflect the opinion of the National Bank of Belgium. We are grateful to participants of the WDN meetings and our colleagues for fruitful discussions.

2 Non-technical summary Based on firm-level data for Belgium over the period , we evaluate the elasticity of labour and average real wage per firm to microeconomic Total Factor Productivity (TFP). We also consider differences between the services sectors and other sectors of economic activity. By doing so we examine wage rigidity at the firm-level and provide microeconomic evidence on the relationship between hours and technology. In a frictionless environment, wages and productivity should go hand in hand. However, wage rigidity and wage insurance may disrupt this relationship. In addition, competition on the labour market may prevent firms to undertake individual wage changes following firm-specific events. Macroeconomic theory shows that when wages are rigid, employment fluctuations account for the adjustment of productivity. Furthermore, wage rigidity also affects the design of optimal monetary policy. Under price stickiness and wage rigidity, pure inflation targeting is no longer optimal. The Central Bank should rather focus on the volatility of both inflation and unemployment. This paper also provides microeconomic evidence on the relationship between hours and technology. The sign of the hours-technology relation is controversial. Some macroeconomic results plead in favour of a negative relationship, consistent with the existence of price stickiness. Some other macroeconomic papers challenge this result. The microeconomic evidence is limited, but suggests a negative within the year impact of TFP on labour. In this paper, we estimate dynamic equations for firm-level average labour cost, employment, hours, wage bill per hour and hours per employee. Our models include firm-specific TFP and other explanatory variables relevant for the determination of labour demand and wages. More specifically, we control for labour force composition, institutional factors related to bargaining practices, sector-level business conditions, as well as firm characteristics. We estimate them using the System GMM technique. The estimates are based on firm-level data obtained from annual accounts and social balance sheet in Belgium. The dataset provides firm-level information on the wage bill, the number of employees and total hours worked Our results can be summarized as follows. First, we find a relatively low elasticity of wages to TFP at the firm-level. It is one of the lowest estimates reported in the literature for other European countries. If we account for variation in hours worked, the contemporaneous sensitivity of wages is only This paper thus points to strong real wage rigidity at the microeconomic level in the sense that there is little room for firms to adjust their average wage to firm-specific changes in TFP. We argue that the dominant role of sector-level collective bargaining in the Belgian wage-setting process may explain this result. Firms respond much more strongly to sector-specific productivity developments and we show that part of the sector-specific productivity developments are transmitted to wage changes through the sector-level wage agreements. Second, we provide microeconomic evidence that hours worked respond positively to technological changes within the year in our data set. This is in contrast to recent findings reported in the literature for Italy and Sweden. We run several robustness tests (among others accounting for variable utilisation of production factors or considering TFP shocks rather than the level of TFP) 1

3 and we find our results robust to the specification tests considered. The hours worked per firm respond positively not only to firm-specific productivity but also to common sector-level technological changes. Third, both wages and employment are less sensitive to TFP in services sector. This result supports the view that the lower frequency of observed price changes for services may be (partly) due to a low response of wages to microeconomic shocks. Indeed, we report evidence of wage sluggishness at the microeconomic level. Further, wages represents a larger fraction of total costs in services sectors than in the manufacturing industries. Lastly, we find that wages are more rigid in the services firms than in the rest of the economy. 2

4 1. Introduction This paper examines the response of the components of firms' total labour costs to technological changes at the microeconomic level. More specifically, we investigate the impact of microeconomic productivity on firms' average real wage bill per employee, firms' employment and hours worked, as well as real wage per hour and hours per employee. We evaluate and compare the elasticity of real wages and labour to firm-specific productivity and investigate the relationship between hours worked and technological change. Introductory courses in microeconomics teach that wages go hand in hand with productivity. In practice, many deviations from the neoclassical textbook model may explain at least temporary deviations from this result. In the medium-run, one would expect a positive relation between productivity and wages. Reasons for deviations from the medium-run relationship include most notably wage rigidity and wage insurance. Firstly, when firms insure workers against wage fluctuations, the response of wages to economic shocks is smoother, and the volatility of hours is larger than under the perfectly flexible case (Boldrin and Horvath (1995)).Guiso et al. (2005), Cardoso and Portela (2005) and Katay (2007) provide microeconomic evidence that firms absorb temporary shocks to productivity. Secondly, downward wage rigidity implies that wage increases are irreversible. Forward-looking firms might be reluctant to increase wages after a positive productivity shock because they will not be able to cut wages in the event of a negative shock. Recent evidence on downward wage rigidity in Belgium that can be found in Dickens et al. (2006, 2007), Du Caju et al. (2007) and Knoppik and Beissinger (2005) points to high downward real wage rigidity in Belgium. Lastly, in a tight and competitive labour market, firms may refrain from wage cuts to avoid that workers leave for better paying companies. Wage rigidity is an important element of the most recent strand of New Keynesian models that give rise to a trade-off between stabilising inflation and output gap after a shock. In the absence of wage rigidity, these models predict that the central bank should fully stabilize inflation at all times and at any cost (Goodfriend and King (1997)). In conjunction with price lumpiness (Christiano et al. (2005)) or real wage rigidity (Blanchard and Galí (2007, 2008)) generate inflation inertia and persistence of output fluctuations. Therefore, following an adverse economic shock the monetary authority must decide whether to accommodate a higher level of inflation or, instead, keep inflation constant but allow for a larger decline in output gap and employment. Pure inflation targeting is no longer the optimal monetary policy, which should rather aim at reducing, but not eliminating, the volatility of both inflation and unemployment. The impact of productivity changes on employment dynamics is controversial. First, under monopolistic competition and flexible prices, a positive technology shock leads to a price reduction, an increase in demand and thereby raises output as well as labour input. Real wage rigidity may exacerbate the response of output and labour. Real wage rigidity impedes labour market adjustment more strongly than nominal rigidity because it implies that real wage adjustment cannot be achieved by freezing nominal wages under positive inflation regime. Models with wage rigidity typically find larger variability of employment in response to productivity shock, as compared to the 3

5 flexible wage scenario (see for example Hall (2005) and Blanchard and Galí (2007, 2008)). 1 Second, under price stickiness, prices and therefore demand remain unchanged. Following a positive productivity shock, the same amount of output is then produced using a smaller amount of labour (provided the shock is not offset by an expansionary monetary policy action for instance). However, firms might also increase output and store unsold goods in expectation of the future price change and increase in demand. A negative relationship between technology shocks and labour may be also explained by a short-run negative impact on production due to necessary adaptation of the stock and/or quality of labour and capital, a low elasticity of demand or high market power. For example, Francis and Ramey (2005) and Smets and Wouters (2007) point out that habit formation in consumption and adjustment costs in investment. may induce a negative relation between hours and technology The empirical evidence both on the macroeconomic and microeconomic level is not conclusive. Starting with Galí (1999), a number of structural VAR analyses point to a negative impact of neutral technology shocks on hours worked. Francis and Ramey (2005) find this result to be robust to alternative VAR specifications and identification schemes. Smets and Wouters (2007) confirm this result in a DSGE model with price and nominal wage stickiness. Basu at al. (2006) find that their growth accounting measure technology has a negative impact on hours within the quarter, but a positive effect after one year. Using both VAR and growth accounting measures of technological changes, Alexius and Carlsson (2007) find that technology shocks are positively correlated with output growth and negatively correlated with changes in hours worked. The finding of Galí (1999) has been challenged on several grounds. It has been argued that the effect of productivity on hours turns positive if one assumes that hours are stationary rather than difference stationary (Christiano et al. (2003, 2004)), if one allows for an investment-specific technology shock (Ravn and Simonelli (2007)), or if one allows shocks other than technology to have long-run effect on labour productivity (Dedola and Neri (2007)). Few recent papers evaluate this question at the microeconomic level. Note that the microeconomic and macroeconomic exercises are not fully comparable because, among others, the former evaluate the impact of idiosyncratic productivity shocks, such as a technological improvement patented by a single firm, while the latter consider the response to a common technological change, such as the introduction of a new software on the market. Based on firmlevel data, Marchetti and Nucci (2005, 2007) and Carlsson and Smedsaas (2007) evaluate the impact of growth accounting productivity measures on total hours worked. These papers suggest that current TFP shock has a negative impact on hours, although the effect of past TFP is positive and compensates for the initial negative effect. The main conclusion of Marchetti and Nucci (2005, 2007) is that the effect of current technology shock on hours is more negative for firms with stickier prices. 1 For example, Hall (2005) extends the Diamond-Mortensen-Pissarides model by considering the case of wage rigidity and shows that wage stickiness may substantially raise and explain the observed volatility of employment and unemployment in response to productivity shocks. Similarly, in the New-Keynesian model of Blanchard and Galí (2007, 2008) real wage rigidity causes fluctuations of output and employment in response to supply shocks and productivity shocks, contrary to the flexible case. 4

6 The aim of this paper is to estimate the sensitivity of firm-level average real labour cost, employment and hours to changes in firm-specific Total Factor Productivity (TFP) based on companies' annual accounts and social balance sheet data for Belgium over the period First, we examine the relative sensitivity of wages and labour to TFP changes. Second, we explore the role of sector-level collective wage agreements in the real wages sluggishness in Belgium. Third, we contribute to the debate on the sign of the relationship between hours and technological change. Fourth, we examine differences between the manufacturing and services sectors and relate them to the differences in price stickiness between the two sectors. Our strategy is to estimate dynamic equations for average wage, employment, total hours worked, as well as hours per worker and average wage per hour. The equations include explanatory variables relevant for labour demand and wage determination and are augmented with TFP measured through the growth accounting framework of Ackerberg et al. (2006). Our results may be summarised as follows: (1) the elasticity of wages with respect to TFP is much lower than the elasticity of labour, (2) compared with microeconomic evidence for other European countries, our estimates of the wage sensitivity to technological changes are in the lower range. Moreover, after accounting for the variation in hours worked, the sensitivity of wages drops further to zero We show that this may be attributed to the fact that the wage dynamics in Belgium is mostly driven by sector-level collective agreements, (3) we provide microeconomic evidence that hours respond positively to technological changes within the year, (4) both wages (per employee) and labour are less responsive to TFP in the services sector than in manufacturing. This corroborates one of the conclusions of the Inflation Persistence Network, namely that the lower frequency of price changes observed for services may be due to wage sluggishness (see Dhyne et al. (2006)). The paper is organised as follows. Section 2 provides a brief overview of the Belgian labour market institutions, it introduces the data and describes the methodology. Section 3 presents our main results and section 4 examines differences between manufacturing and services. Robustness tests with respect to alternative measures of TFP and specifications are discussed in Section 5, while Section 6 concludes. Technical details on the construction of the data set and measurement of TFP shocks are included in Appendix A. Additional results are included in Appendix B. 2. Institutions, data and methodology 2.1 Institutional features of the Belgian labour market In this section, we briefly introduce the main features of the Belgian labour market that are relevant for the interpretation of our results. Notable characteristics of the wage formation process in Belgium include the minimum wage, automatic indexation, a cap on wage increases, and sectoral collective bargaining. As far as employment is concerned, strict employment protection may be eased by early retirement, temporary unemployment, as well as overtime. 5

7 Sector level collective wage bargaining between unions and employers representatives plays a major role in the wage formation process. 2 The wage setting in Belgium may be described as the outcome of three mechanisms. First, a prominent feature of the Belgian labour market is full automatic indexation of nominal gross wages to the so-called health index, which is the consumer price index excluding alcoholic beverages, tobacco and motor fuels. This impedes real wage reductions of job stayers through the pace of inflation. Second, the so called wage norm, set at the national level, is a recommendation for a maximum nominal hourly labour cost increase. It is set by an interprofessional agreement for two years and takes into account, among others, the predicted indexation and evolution of labour costs of the main trading partners of Belgium (namely Germany, France and the Netherlands). Third, sector-level agreements, typically organised separately for white-collar workers and blue-collar workers, specify real wage increases, which often consist of an absolute rise in the minimum pay scale. On top of this some companies have developed firm-level wage bargaining. These features explain why Belgium is characterised as a country with substantial real wage rigidity. However, it should be noted that labour compensation involves extra-wage components such as bonuses, premia and overtime hours, which make total compensation more flexible than the base wage. Employment developments over the last decade have been characterised by changes in the labour force composition. The trends include a reduced proportion of blue-collar workers in private sector employment (from 54% in 1990 to 49% in 1997 and 46% in 2005 according to Social Security statistics), an increasing fraction of part-time workers (part-time workers represent 13.5% of employment in 1990, 16.3% in 1997 and 18.1% in 2005 (OECD (2002, 2004, 2006))), fewer hours worked per employee (the annual number of hours worked per employee fell from in 1999 to in 2005 (OECD (2004, 2006))) and a slightly higher number of employees with fixedterm contracts. Fixed-term contracts represent a small proportion of wage earners in Belgium, 6.3% in 1997 and 8.8% in 2005, in comparison with EU average of 12% in 1997 and 14% in 2005 (Eurostat New Cronos). Among the OECD members states, Belgium has high level of employment protection legislation measured by employment protection of regular workers against individual dismissal, specific requirements for collective dismissals, and regulation of temporary forms of employment (ranking 8th among 28 countries in 2003, see OECD (2004)). On the other hand, flexibility of the labour market is increased by early retirement and temporary unemployment. For firms in distress or restructuring, early retirement is possible under specific conditions for workers aged 50 and more. For short periods, temporary unemployment allows firms to temporarily interrupt, but not breach, labour contracts. Workers then receive unemployment benefit for a defined period and are later re-employed by the same firm under the initial contract terms. Together with changes in the 2 They determine various aspects of compensation, such as pay scales and real wage increases, as well as other aspects, such as training or mobility. Pay scales define a minimum wage by sector and occupation and vary with age or tenure for white-collar workers and some blue-collar workers. Following the EC anti-discrimination rules, relation with age is less common today. 6

8 number of hours (e.g. due to overtime hours), temporary unemployment allows to reduce the number of hours worked, possibly with no change in the number of employees, and can avoid costly layoffs, as does early retirement. 2.2 Data The main variables of interest related to labour costs (total wage bill, number of employees, total hours worked) are taken from firms annual accounts. Almost all firms in Belgium have to file their annual accounts, however, we consider only firms in manufacturing, construction and market services sectors, and we focus on firms with at least 50 employees. We perform a range of consistency checks to identify possible data issues and exclude extreme observations as outliers. Technical details are discussed in Appendix A.1. In our analysis, we estimate equations of employment, labour costs, hours etc. by System GMM. To make sure that sufficient history is available to build lagged instruments, we consider only spells with at least 6 consecutive observations per firm and the variables in levels. Last, we exclude sectors with either too few observations to estimate the production function, from which our measure of TFP is derived, or sectors with production function coefficients substantially different from the income shares. The real wage bill of firm i at time t is denoted as WB it and includes total remuneration and direct social benefits deflated by sector-specific value added prices. Employment, abbreviated as L it, is measured as the average number of employees in full-time equivalent over the year. Average real wage per firm (W it ) is simply obtained as the ratio of the total real wage bill and the average number of employees over the year in full-time equivalent. Total hours worked over the year for each firm are denoted as H it. Value added per sector (VA st ) was obtained from national accounts statistics. Variables related to workforce composition, like the percentage of blue-collar workers (%BC it ), the percentage of women (%WOMEN it ) and the proportion of workers with fixed-length contract (%TEMP it ), are provided in the so called social balance sheet, which forms part of firms' annual accounts since This restricts the sample available to our study to the period The construction of capital stock (K it ) is based on the perpetual inventory method (see Appendix A.2 for details). In what follows, variables in lowercase designate log transformation. We measure average wage per firm as its total labour costs divided by the number of employees in full-time equivalent (as in Katay (2007)), or alternatively by the number of hours worked. This contrasts with empirical papers based on individual wages (such as Biscourp et al. (2005), Cardoso and Portela (2005), Guiso et al. (2005)). These studies focus on job stayers. Analyses that focus on job stayers may underestimate the sensitivity of wages if the wages of job stayers are (partly) set by multi-period contracts. One advantage of our measure is that it also includes employees whose wages might be more easily adjusted than permanent job-stayers, such as new entrants or workers on fixed-term contracts. Evidence that the wage of entrants or movers is more flexible than that of job stayers is provided by Fehr and Goette (2005) and Haefke et al. 3 We disregard the information for year 1996 due to data issues. 7

9 (2007). A potential disadvantage of our measure of wages is that it may vary with changes in the composition of the labour force that are independent of technology changes. We account for this by controlling for the percentage of blue-collar workers, women and workers under fixed-term contract in our equations. Note also that our measure of wages, i.e. firm s average labour cost per employee, may also be more flexible than the base wage because it includes extra-wage components such as overtime hours, bonuses and premiums. In order to account for variations in hours worked (H it ), we also consider the hourly wage, defined as total labour costs over total hours worked. We attempt to capture in our wage equation the impact of sector-level collective agreements of each firm's wages. This is motivated by the considerable importance of sector-level collective agreements in the wage-setting process in Belgium and our estimates confirm their relevance for firms' average wage. The variables are constructed as follows. The nominal index of collectively agreed nominal wage increases at the sector-level for blue-collar workers and white-collar workers respectively, is published by the Ministry of Labour 4 and we deflate it by the corresponding sectorlevel value added deflator to obtain the real measure. We use the logarithm of the real index of collectively agreed wage increases for blue-collar workers and white-collar workers, I B st and I W st, respectively, and multiply these by the percentage of blue-collar workers and white-collar workers for each firm. The measure is not perfect because collectively agreed wage increases are defined at a more detailed level (in terms of sectors, but also occupation and age or tenure). Note that collectively agreed nominal wage increases in Belgium are the result of two mechanisms: indexation and collective agreements concerning real wage increases. We do not attempt to estimate the latter, i.e. we do not try to discriminate between indexation and real wage increases negotiated within sector collective agreements. Rather, we try to evaluate the impact of wage increases induced by sector-level collective agreement that is decided outside the firm on the firm's labour costs. From the point of view of the firm, these have to be compared to the firm's output prices. Therefore, we deflate the collectively agreed nominal wage increases by value added deflator. Discrepancies with respect to the average wage may capture the firm-specific wage policy but also reflect the fact that collective agreements do not apply to more flexible components of labour compensation which include bonuses, premiums and overtime hours paid. Since the aim of our paper is to evaluate the response of wages and labour to productivity, it is crucial to construct unbiased and consistent measures of productivity and avoid spurious correlation with labour. 5 We estimate TFP through the method recently proposed by Ackerberg et 4 Federal Public Service Employment Labour and Social Dialogue (FPS ELSD) 5 Consider the production function for gross output Y it of firm i at time t as a function of labour input, L it, capital stock, K it, and technology index Z it: Y it = F it(l it, K it, Z it). For simplicity, consider the case of constant returns to scale, competitive markets and factor mobility. Profit maximization leads to the following expression for productivity change E[ d ẑit ] = E[dy it - ˆ L dl it - ˆ K dk it], where lower-case variables are measured in logs. Now suppose that the labour coefficient, L, is estimated with an upward bias. Then the estimates of productivity change are also biased E[ d ẑit ] = E[dy it - ˆ L dl it - ˆ K dk it] = dy it - ( L+ )dl it - Kdk it = dz it - dl it, 8

10 al. (2006) who improve on several grounds the estimation procedures of Olley and Pakes (1996) and Levinsohn and Petrin (2003). Ackerberg et al. (2006) point to a colinearity issue that invalidates the estimates by Levinsohn and Petrin (2003). The method of Olley and Pakes (1996) provides unbiased estimates only under relatively strong assumptions. Ackerberg et al. (2006) propose a method that corrects for the colinearity problem and yields consistent estimates of the production function coefficients under a less severe set of theoretical assumptions. It also corrects the estimated TFP for unanticipated shocks and measurement errors. A thorough discussion and technical details are included in Appendix A.3. The data set contains altogether 9621 firm-year observations on 1359 firms with more than 50 employees over the period Table A1 in Appendix A provides more details on the composition of the data set across the sectors considered in the paper. Basic descriptive statistics on the variables introduced above, together with the change in the log of firm-specific productivity ( tfp it ) are given in Table 1. Table 1 - Descriptive statistics Variable obs. mean st dev. P5 median P95 W it (a) L it H/L it WB/H it tfp it I it/k it (b) %BC it %TEMP it %WOMEN it %L>100 it w it l it (h-l) it (wb-h) it va st Notes: Descriptive statistics for firms with more than 50 employees and 6 consecutive annual accounts over the years P5 and P95 refer to the 5 th and 95 th percentile. Lowercase variables are in log. (a) Real average annual gross salary in euro. (b) Investment-capital ratio. 2.3 Specification Equation (1) shows the baseline model that we estimate in Section 3: y it = 1 y it y it tfp it + 2 tfp it k it + 4 va st + 5 (%WHITE it * i W st) + 6 (%BLUE it * i B st) + 7 BLUE it + 8 WOMEN it + 9 TEMP it + 10 L>100 it + i + s + t + it (1) and the term dl it induces a spurious negative correlation between the estimated productivity shock and labour. 9

11 Variables in lower-case are measured in logs and j 's are the coefficients to be estimated. A vector of dummy variables for the 14 sectors considered in the paper are denoted as s, year dummies as t and firm fixed effects as i, "L>100 it " stands for a dummy variable indicating whether the particular firm has more than 100 employees. In equation (1) y it denotes the dependent variable, which can be any of the following variables: wage per worker (w it ), employment (l it ), hours (h it ), hours per worker (h-l it ), and wage per hour (wb-h it ). Hence, we estimate dynamic equations for each component of the wage bill and we also provide estimates for total hours worked and wage bill per hour worked. We include the same set of variables in all equations, which may therefore be viewed as reduced-form equations. We allow for additive sector and year dummies to capture macroeconomic conditions. Robustness tests with respect to sector-specific year dummies, st, are reported in section 5. In addition to sector and year dummies, the choice of variables are motivated by wage determinants typically considered in Mincer-type equations, by institutional wage setting practices, as well as standard labour demand equations. In the wage equation, we control for the composition of the labour force by including the percentage of blue-collar workers, the percentage of women and the percentage of workers with fixed-term contract. We include a dummy variable for firms with more than 100 employees, as well as firm-specific effect to control for unobserved firm characteristics. Because unemployment rates are not available at the sector-level, sector-level business conditions are captured by the log of sector value added. We also include weighted indices of wage increases for blue-collar workers and white-collar workers determined by sector-level collective agreement. Change in the installed capital enters the employment equation and hence it also appears in the reduced form wage equation. Similar equations were used by Katay (2007) for the average wage bill per worker and by Biscourp et al. (2005) and Guiso et al. (2005) for individual wages. 6 Various specifications of labour demand equations have been used in the literature. Assuming that firms are output constrained, employment depends on expected output and the relative capitallabour ratio. If firms are not output constrained and the capital is pre-determined, then firm's employment depends on the pre-determined stock of capital and real wages. Under monopolistic competition, employment also depends on the sector output price (or sector demand). Sector output is then included in the employment equation to capture sector demand. The employment equations becomes dynamic once one takes into account adjustment costs, so we also include the lagged level of employment in the equation. 7 Empirical applications include Nickell and Wadhwani (1991), Arellano and Bond (1991), and Nickell and Nicolitsas (1999). Instead of including real 6 Katay (2007) also estimates similar aggregate wage equation for the average wage bill. The richness of his datasets allows him to take into account additional characteristics, like gender, age, occupation and education. Based on individual wage data, Biscourp et al. (2005) evaluate the asymmetric impact of productivity changes (measured by the growth of value added per employee) on wages after controlling for gender, age, occupation, tenure, firm size, sector of activity and local unemployment rate. Guiso et al. (2005) estimate the permanent and transitory effect of productivity changes on wages, controlling for gender, age, occupation, region, industry and year characteristics. Cardoso and Portela (2005) use the same approach as Guiso et al. (2005) but control only for region, industry and year dummies. 7 We do not consider other types of adjustment costs that lead to more complicate dynamics, such as fixed costs that generate lumpy adjustment and inaction zones. 10

12 wages, we add in the employment equation all variables relevant for the wage determination. This may be viewed as a reduced form approach. In equation (1) we allow for firm-specific fixed effects, as it is common in the literature. This implies that instrumental variables should be used to take into account endogeneity of the lagged dependent variable. The dynamic panel equations are estimated by the System GMM procedure proposed by Arellano and Bover (1995) and Blundell and Bond (1998). We report the two step estimates with standard errors corrected by the Windmeijer (2004) procedure assuming that TFP, firm size, labour force composition, and the impact of sector-level collective agreements on firms' wages are exogenous. Lags of the endogenous variable, capital stock and profits per worker are used as instruments. 3. Results 3.1 Estimating the elasticity of wages, employment and hours to TFP In this section we discuss the estimates of equation (1) for wage, employment, hours worked, wage bill per hour worked, as well as hours per worker. The results are reported in Table 2. Our estimates indicate that the contemporaneous elasticity of wage with respect to productivity is very small, 0.05, while the elasticity of employment is rather large, Adding up the coefficients on current and lagged TFP, the total elasticity of wages is 0.03, which is four times smaller than the elasticity of employment (0.12). The elasticity of total hours worked, which accounts both for changes in hours per worker and changes in the number of employees, is of the same order of magnitude as that of employment. However, the contemporaneous elasticity of hours per worker to TFP equal to This means that firms adjust labour to firm-specific productivity developments mainly through the extensive margin, rather than the intensive margins. It also implies that the rate of utilisation of labour increases following a positive TFP change. 8 Note that we measure wages by firms' labour costs over the number of employees (in full time equivalent). Fluctuations in hours per worker, reported in the social accounts, imply variation in labour compensation. Therefore the estimated response of firm s average wage to productivity changes may capture changes in compensation due to overtime hours or temporary unemployment in addition to the reaction of the wage. In order to verify this conjecture, we estimate a wage equation for labour costs per hour worked. Indeed, the coefficient on TFP is smaller, 0.03, than for wage per worker but significant. 8 Robustness tests with respect to a measure of TFP corrected for variable rate of utilisation are presented in section 5. 11

13 Table 2 - SGMM estimates of equation (1) w it l it h it wb-h it h-l it constant 2.01** *** (0.91) (0.36) (0.34) (0.33) (0.79) dep. var t *** 1.28*** 1.07*** 0.60*** 0.30** (0.11) (0.13) (0.09) (0.10) (0.13) dep. var t *** -0.17** 0.21*** 0.41*** (0.09) (0.12) (0.08) (0.08) (0.08) tfp it 0.05*** 0.22*** 0.22*** 0.03* 0.02* (0.02) (0.05) (0.02) (0.02) (0.01) tfp it *** -0.13*** * (0.02) (0.04) (0.03) (0.01) (0.01) k it *** 0.02*** (0.01) (0.01) (0.00) (0.00) (0.00) va st ** 0.01 (0.04) (0.04) (0.04) (0.03) (0.02) %WHITE it * i W st 0.61*** *** 0.11* (0.1) (0.09) (0.08) (0.08) (0.06) %BLUE it * i B st 0.73*** *** (0.06) (0.07) (0.05) (0.05) (0.04) %BLUE it -0.18*** 0.02* *** -0.03*** (0.05) (0.01) (0.01) (0.03) (0.01) %TEMP it -0.07*** 0.3*** 0.25*** -0.08*** 0.05* (0.03) (0.06) (0.03) (0.02) (0.03) %WOMEN it -0.09** 0.03* ** -0.03*** (0.03) (0.02) (0.01) (0.02) (0.01) L>100 it *** 0.09*** (0.01) (0.03) (0.01) (0.00) (0.00) Sargan p-value (0.07) (0.06) (0.09) (0.34) (0.53) AR(1) p-value (0.00) (0.00) (0.00) (0.00) (0.06) AR(2) p-value (0.64) (0.88) (0.69) (0.18) (0.00) Note: Firms with at least 50 employees and 6 consecutive annual accounts. Two step System GMM estimates are reported with standard errors in parentheses following the correction proposed by Windmeijer (2004). The lagged dependent variable (denoted as dep. var t-1) and the capital stock are treated as endogenous and instrumented with the Arellano-Bond instrument matrix with lags t-4 and earlier, as well as profit per worker. The remaining regressors are treated as exogenous. All equations include separate sector and year dummies but their coefficients are not reported. AR displays the test for serial correlation in the first-differenced residuals. Lowercase variables are in log. The remaining variables are defined in the text. * indicates significance at the 10% level, ** at the 5% level, *** at the 1% level. The coefficients on control variables in Table 2 have the expected sign. Firms with a higher percentage of blue-collar workers and women have significantly lower average wages, all else equal. Also, firms with a higher percentage of workers under fixed-term contract pay ceteris paribus lower average wages. The capital stock has a positive coefficient in the employment equations suggesting complementarities between the two production factors, capital and labour. Value added per sector has a positive and significant impact on wages, implying procyclical wages. The response of employment to value added per sector is not statistically significant. This can be the result of labour hoarding or a strategy to postpone to downturn infrequent support tasks (like machinery maintenance) that are not directly related to output in the period in which they are performed (see Fay and Medoff (1985)). Importantly, the impact of sector-level collectively agreed wage increases at the firm level are positive and significant in the wage equation. The point 12

14 estimates imply that a one percent increase in the collectively agreed wage induces firms to raise wages on average by 0.61 an 0.73 percent for blue-collar and white-collar workers, respectively. The coefficient on white-collar workers is slightly smaller than that of blue-collar workers. One reason may be that a larger fraction of earnings of white-collar workers is not subject to collective agreements, such as premiums and bonuses. Our results may be summarised as follows. First, labour is more sensitive to TFP than wages. This may be the result of significant wage rigidity in Belgium and, as will be discussed below, the importance of sector-level collective agreement in the wage formation process. Second, following a TFP change, firms adjust labour both on the intensive and extensive margins. In the presence of adjustment costs in the short-run, firms may adjust hours worked more easily than the number of employees, for example through overtime hours and temporary unemployment. 9 However adjustment through the extensive margin plays a dominant role. Third, our estimates point to a positive relation between hours and TFP. The following two subsections examine in more detail the finding of a low elasticity of wages and a positive elasticity of hours with respect to TFP. 3.2 The low response of wages to productivity In this section we compare the elasticity of wages in Belgium with estimates reported for other countries and investigate the role played by sector-level collective agreements. Table 3 reports the microeconomic point estimates of the impact of firm-specific productivity changes on wage changes for France, Italy, Portugal and Hungary. Following the methodology of Guiso et al. (2005), a number of authors have decomposed the impact of firm-level productivity on individual or firmlevel wages into one effect due to permanent productivity changes and another effect due to transitory variation in productivity. Taking a different perspective, Biscourp et al. (2005) investigate the asymmetric response of wages to positive and negative productivity changes. The numbers reported in Table 3 are not fully comparable due to different approaches, data and sample definitions, but they still provide a basis for comparison. Below, we consider the main differences. First, Guiso et al. (2005) focus only on the manufacturing sector. Our own estimates indicate that the elasticity of wages to productivity is lower when services are included in the sample (see Section 4). Second, papers based on individual wage data restrict their sample to job stayers. However Fehr and Goette (2005) and Haefke et al. (2007) report that entrants' wages are more flexible than incumbents' wages. 10 Hence, focusing on job stayers may bias the estimated 9 See Fuss (2008) for evidence that variations in hours per worker and the number of days worked are significantly lower in case of sales declines and wage bill contractions. 10 Fehr and Goette (2005) provide evidence that the wages of job movers are significantly less rigid than wages of job stayers. From the Swiss Labour Force Survey over 1996, 1997, 1998, they find that workers experience a wage cut as soon as their notional wage cut reaches 10 percent for job movers, 20 et 25 percent for part-time job stayers and 30 to 40 percent for full-time job stayers. Haefke et al. (2007) estimate the elasticity of the aggregate wage to labour productivity for the US. Over the period , the elasticity varies between 0.64 and 0.94 for newly hired workers out of non employment, against between 0.17 and 0.37 for all workers. 13

15 elasticity downwards. Third, the measure of wage may influence the results. Among others, our results confirm that considering wage per worker or wage per hour makes a difference. Considering significant coefficients, the total impact of TFP on wage per hour in Belgium amounts to 0.05, close to the estimated impact for France (0.04). It is lower than the estimated permanent effects of productivity obtained for Italy, 0.07, Portugal, 0.09 and especially the sum of the permanent and temporary effects in Hungary, Keeping in mind the limited comparability of the results, two observations emerge clearly. First, the estimated microeconomic elasticity of wages to productivity is far below one. Second, the elasticity is particularly low in Belgium. Concerning the first point, two potential explanations were brought up in the literature: wage insurance and wage rigidity. The first is based on the idea that risk neutral firms insulate wages of risk-averse workers from adverse shocks to production. For example, in a contract model where firms insure workers against income fluctuations, Boldrin and Horvath (1995) show that the response of wages to shocks is smoother, and the volatility of hours worked is larger than in a situation without wage insurance. As a result, the correlation between output and wages is lower (and possibly zero) as compared to flexible wage environment, while the correlation between output and hours is larger. Another explanation may be derived from the evidence on downward wage rigidity, measured either from the asymmetry of the wage distribution (see the recent work of Dickens et al. 2006, 2007) or as the inability of wages to respond to negative shocks (as in Biscourp et al. (2005)). The sensitivity of wages to technology changes under downward rigidity can be reduced both downwards and upwards because firms may be reluctant to raise wage following a positive productivity shock that would be difficult to revert in case of future adverse economic conditions. Lastly, labour market competition and efficiency wage considerations may explain the low response of firms' wages to firm-specific shocks. For example, in a tight labour market it may not be desirable for a company to reduce wages following a negative productivity shocks, because it makes other companies more attractive for its workers. Further this may generate adverse selection problems. This argument would explain why the firms' wage response to firm-level shocks is low, while the elasticity to common shocks is large, as shown in Table 4 below. These arguments may explain why the sensitivity of wages to productivity is low. But it does not explain differences across countries, unless there were large differences in the degree of wage insurance or wage rigidity. Several reasons may explain differences between the results for Hungary and other Western European countries. In the case of Hungary, the response of wages to productivity is not only the largest but also there is no evidence of full insurance against temporary fluctuations in productivity, contrary to the results reported for Italy and Portugal. Katay (2007) points to the fact that Hungary is an economy in transition that experiences large macroeconomic shocks. Less developed financial markets imply that firms have fewer financial instruments to provide insurance. He also argues that households have a strong preference for present consumption, and thereby a lower preference for smooth wage contracts. In addition, according to OECD (2004), Hungary is among the countries with the most flexible labour market. 14

16 Table 3 - Survey of microeconomic estimates of the sensitivity of wages to productivity country wage measure sample Katay (2007) Hungary, firm's average net private sector real earnings of full-time workers Guiso et al. Italy, individual (2005) manufacturing earnings of fulltime job stayers Cardoso and Portugal, individual real Portugal private sector gross hourly (2005) earnings Biscourp et al. France, individual net (2005) private sector hourly earnings of full-time job stayers Fuss and Belgium, firm's average Wintr (2008) private sector real wage bill firm's average real wage per hour Note: denotes that the coefficient is insignificant. productivity measure permanent effect transitory effect TFP (Levinsohn and Petrin) value added / employee sales value added / employee TFP (Ackerberg et al. (2006) positive effect negative effect total effect current: lagged: current: lagged: 0.00 current: lagged: current: 0.05 lagged: current: 0.03 lagged:

17 Company and plant-level agreements are the dominant form of wage bargaining, with no coordination by upper-level association and no centralisation. In contrast, industry level agreements are common in the remaining European countries, combined with firm-level agreements (in France and Italy), or with central-level agreements (in Portugal and Belgium), and a medium or high level of coordination. The coverage of collective agreements is much higher in Belgium, France, Italy and Portugal (around 90%) than in Hungary (36%). 11 In sum, because in Hungary wage bargaining takes place on the firm level with little or no co-ordination, firms are have more freedom to change wages in response to TFP shocks. Differences between Western European countries are hard to relate to indicators of coverage, centralisation and co-ordination of wage bargaining. Focusing on Belgium, first note that the low elasticity of real wages we find is in line with previous findings of high downward real wage rigidity in Belgium (see Dickens et al. (2006, 2007) and Du Caju et al (2007)). In addition, it is consistent with the predominant role played by sector-level collective agreements. Table 2 reports estimates of the response of wages to firm-specific productivity after controlling for common macroeconomic evolution, and more importantly, sector-level collective agreements. Even though firms' average wages do not respond strongly to the idiosyncratic (firm-specific) component of TFP in our wage equation, they might respond much more vigorously to the sectorspecific TFP. Such situation could arise if productivity developments at the level of sectors were captured by sector-level agreements, and if the impact of sector-level bargaining on firms average wages (%WHITE it * i W st and %BLUE it * i B st) was driven mostly by sector agreements, i W st and i B st, rather than by the firm workforce composition. In order to verify this conjecture, Table 4 reports System GMM estimates of the wage equation for alternative specifications. Column (1) reprints the estimates from Table 2 for comparison. In column (2), we replace additive sector and year dummies by multiplicative dummies. Once sector-specific time dummies are allowed for the coefficients on sector-level collectively agreed wage increases in the wage equation, %WHITE it * i W st and %BLUE it * i B st, become smaller and insignificant. This suggests that these variables mainly capture differences across sectors rather than differences across firms. 12 Second, in order to evaluate the extent to which sectorspecific developments in productivity impact on average wage, column (3) reports the sectorspecific component of TFP, tfp st, 13 in addition to the idiosyncratic component, tfp it, omitting sectorlevel collectively agreed wage increases. The total impact of sector-specific component of TFP amounts to 0.27 and is much stronger than the impact of the firm-specific TFP. 11 The measure shows the percentage of employees covered by collective agreements (i.e. it includes not only union members but also other employees to whom the union-negotiated contact applies). Data refer to Source: Traxler and Behrens (2002). 12 The result is driven by the fact that wage increases agreed in sector-level collective agreement follow the same trend for blue-collar and white-collar workers of the same sector but different patterns across sectors. 13 tfp is obtained as as z it = y it - ˆ L l it - ˆ K k it, where ˆ L and ˆ K are production function estimates obtained by the Ackerberg et al. (2006) procedure (see Appendix A.3). tfp it is then obtained as the residual from a regression of z it on sector specific time dummies. The sector specific time dummies constitute the sectorspecific component of tfp, i.e. tfp st. 16