Productivity differentials and firm strategy in Italian manufacturing: some evidence Enrico Tundis Graduate School of Social Sciences, University of Trento Enrico Zaninotto Department of Computer and Management Sciences, Univerity of Trento Preliminary draft June 1, 2012 Abstract This paper studies the dynamics of the productivity of a sample of Italian manufacturing firms over the period 1996-2006. The breakdown of productivity growth into a technlogical and an efficiency component permits to gauge clearer evidence of the increasing dispersion of firm performance over time, with some firms that achieve technological advances and move the best-practices frontier, and other firms that apparently seem to not move from their positions and get further away from the frontier. Moreover, the analysis suggests how observed productivity heterogeneity may be eventually related to different firm strategies, mostly associated to different management of workforce. The first evidence put forward reinforces the idea of a dualism between two groups of firms that compete either through innovation or cost, particularly labour cost. 1. Introduction Earlier studies on the economic slowdown that has characterized Italy since mid nineties gave a pessimistic view about the future of the Italian productive system. In the face of new challenges, manufacturing and perhaps the entire economy seemed inert and to move towards its decline (see Faini, 2004 and Onida, 2004 among others). Recent studies, however, suggest a more complex picture. The competitive environment faced by Italian manufacturing has changed radically since the mid nineties due principally to the emergence of countries with low labour cost (especially China) and to the process of European integration, culminating with the introduction of the euro (Brandolini and Bugamelli, 2009). One hypothesis put forward is that the adoption of the euro, by eliminating competitive devaluation often used in the past, has had a profound impact on the Italian productive system populated by firms with very different characteristics, independently on the geographical location and sectoral specialization. These heterogeneous firms would react in different ways to external to the firm changes: some firms were able to face the new “competitive game”, while others went through a low growth path. In this regard, Dosi et al. (2011) suggest “Turbulence underneath the big calm”. By analyzing a sample of over 100,000 firms in all economic sectors, they found an increase of the dispersion of labor productivity between 1989 and 2004, giving rise to a kind of neo-dualism among firms. Bugamelli, Schivardi and Zizza (2010) reach similar conclusions. Looking at labour productivity as indicator of firm restructuring they argue that the euro adoption would not affect so much the reallocation of firms across sectors, but had principally a within firm effect, forcing firms to internal changes. Boeri and Garibaldi (2007) argue instead that the origin of the decline in labour productivity and the increase of its dispersion is relate to the labor reforms and the establishment of a stratified model of labour market, with a rigid and a flexible component. This study fits in with this strand of research with two contributions. First, the study looks at total factor productivtity instead of labour productivity. Specifically, by using the Malmquist index and its decomposition it considers productivity dynamics as stemming from two different mechanisms: how much closer or farther away firms gets to the best-practice frontier and the extent to which the best-practice firms are changing (improving or deteriorating) their positions in the input-output space over time, allowing a comparison among the evaluated firms. The results that emerge give further confirmation of the underlying heterogeneity of firm behaviour. In particular, the first part of the study shows a discontinuity of behaviors after the adoption ot the euro, highlighting the gap between firms that contribute to technological advancement, and firms that worsen their performance and move away from the frontier. Moreover, after classifying firms into categories based on their productivity dynamics, the second part of the study explores the relationship between firm belonging to one category and a set of firm characteristics, in particular the cost and composition of workforce, and financial constraints. Firms belonging to less dynamic group, i.e. those firms characterized by slower productivity growth, appear to use labour of lesser quality and draw from flexible labor market more extensively then other firms. Although preliminary, the results are consistent with the idea that some firms have taken advantage of the emergence of two tier labour market (Boeri and Garibaldi, 2007). The paper is organized as follows: section 2 presents the productivity estimation strategy; section 3 decribes the database; section 4 presents the results emerging from the analysis of productivity growth and its decomposition. Finally, section 5 tries to isolate some factors that characterize different classes of performance. 2. Measuring and decomposing Total Factor Productivity This paper uses a non-parametric approach in the frontier framework as estimation strategy. Specifically, it is employed the DEA approach (Cooper et al, 1978; Banker et al., 1984) since it shows some caracteristics that make this technique very appealing for the analysis. First, DEA yields estimates of efficiency and an estimator of the production frontier (Kneip et al., 1998, 2008; Simar and Wilson, 2008). Second, when !rms are likely to employ different technologies, DEA estimates are among the most robust (Van Biesebroeck, 2007). Besides, this technique does not make a priori assumptions on the shape of frontier function and establishes a best-practice frontier among the observed firms based on direct comparison process. Finally, in the DEA framework it is possible to obtain measure of Total Factor Productivity change by means of the Malmquist index (Fare et al., 1992, 1994). This index computes firm productivity change over time directly from input and output data and allows meaningful decompositions of productivity dynamics between a technological 2 component – related to best-practices frontier shift – and a component linked to efficiency improvements – related to firms distance from the frontier. In this frontier framework, the shape of the frontier and consequently the estimated productivity depends on the choice of the production set. One possibility is to employ contemporaneous frontiers, i.e., production sets are constructed at each point in time from the observations at that time only. In this case, production sets can expand or contract from one year to another and outward – technical progress – as well as inward – technical regress – shift of the frontier can occur with respect to the base time period considered. However, a single intertemporal production set by using the full dataset, or a sequential frontier by using accumulate data until the baseline year (Tulkens and Vanden Eeckaut, 1995) can be alternatively constructed. In both cases inward shift of the frontier is not allowed. As in the the original Malmquist TFP approach (Färe et al., 1992, 1994) and the majority of Malmquist TFP application in the relevant literature (Heshmati, 2003), this study employs the contemporaneous frontier approach allowing for upward as well as inward shift of the frontier. Indeed, this choice is particularly useful since it allows to observe how best-practices firms and firms below the frontier move with respect to each other. 2.1. The Malmquist index Consider a firm producing a vector of outputs, y ! "+M , from a vector of inputs, x ! "+S . Assume a convex production possibility set with freely disposable inputs and outputs. Then, the output distance function1 can be de!ned on the technology T = {( x, y ) : x!can produce y} as (Shepard, 1970): ( + ! y$ D ( x, y ) = inf )! > 0 : # x, & ' T , ! * " !% - (1) This distance function is relative to each firm and can be interpreted as the potential increase of output that can be achieved by the firm that uses a given amount of inputs. In particular the scalar " # (0,1] identify the potential expansion of the output y such that the production possibility ( x, y " ) lies on the efficient frontier T at time. Therefore a firm will be efficient at time t (lay on the frontier) iff D ( x, y ) = 1 . Since ! the production possibilities T in (1) is not known, it must be estimate it in order to obtain estimates of the distance. To overcome this issue, we use the Data ! Envelopment Analysis estimator. The DEA production set assuming Constant Return to Scale (CRS) (Cooper et al, 1978) can be described by: N N %' )' T̂CRS = &( x, y ) : " ! j y jm ! ym ,!!m = 1,..., M;!" ! j x js ! xs ,!s = 1,..., S;! ! # $+N * (' +' j=1 j=1 (2) and assuming Variable Return to Scale (VRS) by (Banker et al 1984): 1 An input-oriented distance function can be simmetrically defined. In this paper we present however only the output-oriented case. 3 N N N %' )' T̂VRS = &( x, y) : " ! j y jm ! ym ,!!m =1,..., M;!" ! j x js ! xs ,!s =1,..., S;!" ! j =1,!! # $+N * (3) '( '+ j=1 j=1 j=1 where in both cases T̂ is an estimate based on the observed data of the true production set T. Consistent estimators of D ( x, y ) defined in (1) can then be obtained by substituting the true, but unknown, production set T whit the estimator T̂ (Simar and Wilson, 2008). As practical matter, estimates of D ( x, y ) , assuming CRS, can be computed by solving a linear program. Specifically, the distance of a firm from the empirical production frontier, is estimated by solving the following linear programming model: #N ' p k %" " j y jm ! ! yim ,!!m = 1,..., M, % % j=1 % 1 = max $ ( p, k = t, t + *t N ! D̂pCRS ( x k , y k ) % % p k %" " j x js ! xis ,!!!!!!s = 1,..., S % & j=1 ) (4) where D̂pCRS ( x k , y k ) is the estimated distance of a firm at time k = t, t+!t from the CRS frontier at time p = t, t+!t. Estimates of distance assuming VRS can be computed (Banker et al., 1984): #N ' p k ! ! yim ,!!m = 1,..., M, % %" " j y jm % j=1 % %N % % % 1 p k = max $" " j x js ! xis ,!!!!!!s = 1,..., S ( p, k = t, t + *t VRS ! D̂p ( x k , y k ) % j=1 % %N % %" " j = 1 % %& j=1 %) (5) For each firm, the Malmquist index represents productivity changes between two period, t and t+!t. This index can be derived as the ratio of distances from the CRS production frontier – composed by the best-practice in the observed set of firms – in each period. The link between the calculated distances and TFP change is as follows: Malmquistt = !TF̂Pt = D̂tCRS ( x t+!t , y t+!t ) D̂tCRS ( x t , y t ) (6) This is the ratio between the distance of the firm in period t+!t from the frontier in period t and the distance in period t from the frontier in period t+!t. It is also possible to define the Malmquist index with respect to the frontier at time t+!t as follow: 4 Malmquistt+!t = !TF̂Pt+!t = CRS D̂t+!t ( x t+!t , yt+!t ) (7) CRS D̂t+!t (xt , yt ) Fare et al. (1994) defined a Malmquist index as geometric average between the two indexes defined in (4) and in (5) as follows: 1 " D̂ CRS ( x , y ) D̂ CRS ( x , y ) %2 Malm = $ t CRS t+!t t+!t ! t+!tCRS t+!t t+!t ' D̂t+!t ( x t , y t ) '& $# D̂t ( x t , y t ) (8) The index can be decomposed as follows: 1 D̂ CRS ( x , y ) # D̂ CRS ( x , y ) D̂ CRS ( x , y ) &2 Malm = t+!tCRS t+!t t+!t " % tCRS t+!t t+!t ! tCRS t t ( D̂t ( x t , y t ) %$ D̂t+!t ( x t+!t , y t+!t ) D̂t+!t ( x t , y t ) (' !!!!!!!!!!=!EffCh "!TeCh (9) where productivity change is splitted into two parts. The first part is the ratio of the distances of a firm from the frontier at two different time and shows how much closer (or farther away) a firm gets to the best-practice frontier. It will be higher (lower) than unity if there has been an increase (decrease) in ef!ciency. The second part can be considered as a proxy of the shifts in the empirical production frontier (i.e., the growth rate of technological progress), from t to t+!t and points out to the extent to which the best-practice firms are changing their performance (improving or deteriorating), allowing a comparison to the evaluated firms. Efficiency change (EffCh) can be further decomposed as follows (Fare et al. 1994): VRS CRS D̂t+!t x t+!t , y t+!t ) # D̂t+!t x t+!t , y t+!t ) D̂tVRS ( x t , y t ) & ( ( (( EffCh = " %% VRS ! CRS D̂tVRS ( x t , y t ) $ D̂t+!t ( x t+!t , y t+!t ) D̂t ( x t , y t ) ' !!!!!!!!!!!!!!!=!PEffCh " SEffCh (10) where PEffCh and SEffCh are measures of pure efficiency change – efficiency change with respect to the VRS frontier – and change in scale efficiency, respectively. Values higher (lower) than unity indicate an increase (decrease) of the related quantities. While pure efficiency change and scale efficiency change are related to Variable Return to Scale (VRS) frontiers movements between two different periods, TeCh variation still refers only to Constant Returns to Scale (CRS) frontier shifts over time. Wheelock and Wilson (1999) observed that if a generic firm in the input-output space remains fixed between time t and t+!t, and the only change that happens is in the VRS estimate of technology, the TeCh component, as measured in previous equations, will be equal to unity indicating no change in technology – since the only way for TeCh to change is if the CRS estimate of the technology changes. Such being the case, the CRS estimate of the technology is then statistically inconsistent. Since the VRS estimator is always consistent (Kneip et al., 1998), a further decomposition of Technological change is proposed by introducing also VRS estimates: 5 1 " D̂VRS ( x , y ) D̂VRS ( x , y ) %2 TeCh = $ tVRS t+!t t+!t tVRS t t ' ( $# D̂t+!t ( x t+!t , y t+!t ) D̂t+!t ( x t , y t ) '& 1 " D̂CRS ( x , y ) D̂VRS ( x , y ) D̂CRS ( x , y ) D̂VRS ( x , y ) %2 t t+!t t+!t t t t ' (11) !!!!!!!!!!!!!!!!(!$ tCRS t+!t t+!t ! tCRS t t VRS VRS $# D̂t+!t ( x t+!t , y t+!t ) D̂t+!t ( x t+!t , y t+!t ) D̂t+!t ( x t , y t ) D̂t+!t ( x t , y t ) '& !!!!!!!!!!=!PTeCh ( STeCh where TeCh is further decomposed into Pure Technical Changes – PTeCh – and Scale Technical Changes, i.e. changes in the scale of technology – STeCh. The first component is the geometric mean of two ratios that measure the shift in the VRS frontier estimate relative to the firm’s position at times t and t+!t. When PTeCh is greater than unity, it indicates an expansion in pure technology, i.e., an upward shift of the VRS estimate of the technology. STeCh provides information regarding the shape of the technology. It describes the change in returns to scale of the VRS technology estimate at two fixed points, which are the firm’s locations at times t and t+!t. When STeCh is greater than unity, this indicates that the technology is moving farther from CRS and the technology is becoming more and more convex. On the contrary, when this index is less than unity it suggests that the technology is moving toward CRS; the index equal to unity suggests no changes. Besides, different decomposition of technological change is possible. Technical progress, in fact, can be independent or dependent on the change of the composition of input used and/or output produced by the firms. Therefore, technical change component can be rewritten as follows (Fare and Grosskopf, 1996): 1 # D̂CRS ( x , y ) D̂CRS ( x , y ) &2 t+!t t ( " TeCh = CRS " % tCRS t+!t t+!t t+!t CRS D̂t+!t ( x t+!t , y t+!t ) %$ D̂t+!t ( x t+!t , y t+!t ) D̂t ( x t+!t , y t ) (' D̂tCRS ( x t , y t ) 1 # D̂CRS ( x , y ) D̂CRS ( x , y ) &2 t t t t+!t t ( !!!!!!!!!!!!!!!!"!% t+!t CRS CRS %$ D̂t ( x t , y t ) D̂t+!t ( x t+!t , y t ) (' !!!!!!!!!!=!MaTeCh " ObTeCh " IbTeCh (12) MaTeCh (Magnitude Tchnical Change) is related to Hicks-neutral technical change. If the magnitude effect is greater (lower) than unity, it means that output of the same composition but greater (lower) in terms of volume is obtained with the same input mix. IbTeCh (Input Biased Technical Change) refers to a non-neutral shift in the bestpractice production frontier due to different input mix, while ObTeCh (Output Biased Technical Change) refers to a non-neutral shift in the best-practice production frontier due to different output mix. Values of ObTeCh (or IbTeCh) greater than unity indicate that the biased technical change amplifies the TFP growth, and values of ObTeCh (or IbTeCh) less than one signify that the biased technical change shrinks the TFP growth. 6 3. Data and descriptive analysis The primary source of the data is the Bureau Van Dijk’s AIDA database, which provides detailed information on the financials, geographical localization, number of employees and local units for a large sample of limited liability Italian firms. From the original collection of data it was selected a sub-sample of single-location manufacturing firms, which were continuously active during the period 1996-2006. Since the original employment figures were missing for several firms, data were supplemented with information on workforce from the National Institute of Social Security (INPS). From this additional source it is obtained the yearly average number of employees for all firms in the sample. Moreover the data allow to decompose workforce into white and blue collars as well as between full and part-time contracts for the eleven years covered in this analysis. In the end, the empirical analysis exploits an original dataset containing information on 7,712 (84.832 observation) Italian manufacturing firms over the period 1996-2006. The database represents a unique collection of data for Italy and allows to deepen the understanding of the dynamics of incumbent firms on a relatively long period of time. Moreover, by dealing with single-location firms it allows to work at a level of analysis that is as close as possible to the single establishment level. Furthermore, focusing on single-location firms, changes such as mergers, acquisitions and divestitures affect only marginally the group of firms in the sample. Besides, it is neutralized the spurious effect stemming from the intra-group reallocation of equipment and personnel. The industry distribution of our data set generally reflects the distribution of firms descripted by the ISTAT “8° Censimento Industria e Servizi” in 2001 – the mid point in the observation period (See Table A1 in Appendix A). 3.1. Input and output variables Input and output variables are constructed from balance sheet data, with the exception of data on labour. The raw data were correct and deflated in order to obtain real values. In this study we have used sectoral deflators constructed by using ISTAT data. Output is measured by the amount of revenues from sales and services at the end of the year, net of inventory changes and changes to contract work in progress; labor input is measured as the total number of employees at the end of the year. Two intermediate inputs are considered: (a) costs of raw materials consumed and goods for resale (net of changes in inventories) and (b) cost of services; capital stock estimate in a given year is proxied by applying a modified perpetual inventory method on nominal value of tangible !xed assets over the period analysed (See Appendix C for further details on the estimation procedure). All monetary measures are expressed in thousands of euros and have been de"ated by the proper industry level index. The deflator for the turnover variable was constructed by processing the time series of national production. The deflator for intermediate inputs is constructed using a weighted deflators of production, with weights calculated as the average of the column coefficients of input/output matrix for the year 2001 of a set of Italian regions. Table A2 in appendix A shows descriptive statistics on the variables of input and output for 2006, the final year of observation in our data set. 7 3.2. Outliers treatment Several authors addressed the problem of the presence of outliers in nonparametric efficiency estimation (Wilson, 1993, 1995; Simar, 2003; Banker and Chang, 2006). In fact, DEA produces efficiency scores by comparing the input/output combination of each firm with respect to a piecewise linear frontier obtained as convex combinations of the best performing firms in the set. This implies that measurement errors for those observations defining the frontier could cause distortions in the measured efficiency for the entire population. In order to detect outliers, we carried a !rst analysis to check the impact of each single observation on the distances of the nearest !rms – whose distance depended from that particular observation – using a method based on the concept of leverage, that is the effect produced on the efficiencies of all the other firms when the observed firm is removed from the dataset (Stosic and Sampaio de Souza, 2005). Afterwars, observations with a wider impact on nearest !rms are discarded from the final calculation (see Appendix B for detail on outlier detection procedure). 4. The dynamics of productivity The distances from the estimated industry frontier at 2-digit level are calculated for each firm and combined to construct and decompose Malmquist indexes. Specifically, we calculate annual growth rate of quantities of interest for each firm and afterwards we take weighted average of annual growth rates for each industry, rather than estimates related to individual firm, accounting for the relative importance of each observation whose productivity index is entering into the average (Zelenyuk, 2006). As first exercise, we compare the annual gowth rates of productivity among industries for the entire period 1996-2006 (Table 1). Some observations are immediate. Productivity, with the exception of few industries such as Machinery equipment and Electronics, shows annual growth rates below 1% everywere. In the majority of industries, the annual growth rate is less than 0.5% and in some industries is even negative. Furthermore, the two components – efficiency change and technological change – behave differently in determining productivity trends (Table A4 in the appendix A). There are industries, such as Textile and Wearing, Leather, Metal products, and Electronics, where, on average, technology improves and efficiency declines, while there are others, such as Paper and Chemical industry, where the opposite occurs. Finally there are industries in which both components show positive sign, as in the Machinery and Transportation equipments. Then, the whole period is splitted into three subperiods: 1996-2000, 2000-2003, 2003-2006. This breakdown allows to capture three different phases: the entrance into the euro area (1999), the negative cycle centered on 2001, and the period before the current crisis, respectively. Two facts emerge. First, productivity shows a negative trend in all industries in the first period (Table 1). On the contrary, the central and final periods show a general increase of productivity. Second, the sign of the two components of productivity growth changes radically among subperiods (Figure 1). 8 Table 1. Malmquist index and its decomposition into Efficiency Change (EffCh) and Technological change (TeCh). Entire period and subperiods. Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipment Electronics Tansportation equipment Other manufacturing Subperiod 1996-2006 1.069 0.353 -0.004 0.268 0.090 -1.004 -0.032 0.490 0.068 0.273 1.019 1.224 0.558 0.196 1996-2000 2000-2003 2003-2006 -0.849 -0.355 -1.001 -0.068 -1.579 -4.327 -1.223 0.041 -0.702 -0.218 -0.510 0.500 -1.010 -0.628 1.624 0.823 0.860 0.975 1.908 8.473 1.794 1.738 0.391 1.239 0.767 1.200 0.523 0.055 6.992 1.007 1.295 0.479 0.765 -4.522 -0.164 0.050 1.144 0.346 3.544 2.652 2.631 1.649 In the first subperiod, uniformly in all sectors, there is a sort of “technological regression” (negative sign of TeCh). The downward frontier shift is however partly offset by efficiency gains. We interpret this evidence as a general inability of bestpractices firms to reach at the end of the period the best production combinations formerly obtained. Moreover, this could be indicative of the pervasive difficulty of the entire productive system associated to increased competitive pressure arising from the introduction of the euro. Under these circumstances, firms seek a recovery of competitiveness in terms of efficiency, being on average closer to the frontier. Shifting the focus to the other subperiods, productivity grows in all industries driven primarily by technological advancement – indicated by the outward shift of the frontier that in some cases is strengthened by improved efficiency – during the second subperiod2. Clearer is the trend that occurs in subperiod 2003-2006, the exit from the 2001 crisis, where technological change keeps on to be positive while everage efficiency decreases. This evidence suggests a process of technological advancement that involves only part of the observed firms, principally those closer to the frontier, while part of firms lags behind. Figure 2 clarifies this dynamic plotting the evolution of the probability density of estimated efficiencies, i.e. distances from the frontier, in the particular case of Electronics industry in the first and last subperiods respectively. At the end of the first subperiod (year 2000) efficiency distribution has shifted to the right indicating a generalized increase of average efficiency. The evolution changes drastically in the second subperiod (2003-2006), in which the leftward shift of the distribution indicates a widespread loss of efficiency for many firms. 2 The coincidence with the cycle should be considered with caution, given the particular sensitivity of observations to start and end points of the cycle. 9 $ ,%3)(792*'(C ,B/(.'*)9+ ,19/*C<297 ,A2*!>)7/3 ;/<)9, ,?=@@)9 # -.)>'(/3, ,1)67'3) ,:225 ,%3)(792*'(C ,D7.)9,>/*E ,:225 ,?=@@)9 ,B)7/3 ,8)/7.)9 -.)>'(/3, ,B/(.'*)9+ ,19/*C<297 ,4225 ,A2*!>)7/3 !! 1)(.*2320'(/3,-./*0) ! ,D7.)9,>/*E ,8)/7.)9 ,1)67'3) ,B)7/3 s ,;/<)9 !$ ,;)7923)=> ,;)7923)=> "FFG!!### !! !" !##H!!##G # " ! %&&'(')*(+,-./*0) 5 5 Figure 1. Technological Change (TeCh) and Efficiency Change (EffCh). Subperiods: 1996-2000 and 2003-2006. 2006 3 Density 2 1 0 0 1 2 Density 3 4 2003 2000 4 1996 0.5 0.6 0.7 0.8 Efficiency 0.9 1.0 0.5 0.6 0.7 0.8 Efficiency 0.9 1.0 Figure 2. Kernel density estimation of efficiency in the Electronics industry. Our results are in line with other studies. Aiello, Pupo and Ricotta (2009) found increases of productivity starting from 2003, after the fall occurred between 1996 and 2003. Moreover, after the introduction of the euro, there is evidence of a profound process of firm restructuring in the Italian manufacturing in terms of technological content, quality, etc. (Rossi, 2006; Brandolini and Bugamelli, 2009; Bugamelli, Schivardi and Zizza, 2010; De Nardis, 2010) that support the evidence of an increase of productivity dispersion. An increase of labour productivity dispersion of firms is also recognized in Dosi et al. (2011) mostly before 2000. The estimation strategy employed enables to deepen the analysis by isolating the effect of scale and input composition on productivity change. The evidence is that 10 increases of efficiency in the first subperiod have occurred simultaneously through a reduction of pure inefficiencies and a recovery of scale (Table A5 in the appendix A). Conversely, similar trends are not so clear and pervasive afterwards. This gives additional evidence of two different stories between what took place at the beginning of the observed period and what happened subsequently. The evolution of average size of firms further support this idea. The average number of employees increased between 1996 and 2000, again substantially in all sectors, but subsequently stabilized (Table A6 in the appendix A). The average size expressed in terms of nominal turnover follows the same pattern, although less evident, till 2003 and then grows in the last period. Looking at the best-practices frontier, it is evident how in the first period the two components – pure and scale technological change – determine the downward shift of best-practices frontier in different manner depending on the industry, while in the last subperiod a general upward shift of the frontier pravails (Table A7 in the appendix A). At the same time, frontier generally tends to modify its shape with a movement towards the constant return to scale (negative sign of growth rate of STeCh). Further insighs come from the analysis of the nature of technological progress (table A8 in the appendix A). The apparent technological regression observed during the first subperiod is principally due to Hicks-neutral shift. This fact reinforces the idea that the downward shift of the frontier is the effect of a general reduction of the use of production factors due to a crisis of competitiveness. On the contrary, the outward shift of the frontier in subsequent periods is both Hicks-neutral and input biased. A plausible interpretation of this evidence is that during the years around the crisis of 2001 some firms have undertaken a process of restructuring with gains of competitiveness and better use of production factors. It would be also the effect of an increased production flexibility due to the introduction of ICT. 5. Productivity and firm strategy The evidence put forward above suggests a generalized attempt to reduce inefficiency as first reaction of firms to increased competition. Afterwards, some firms significantly have changed their structure in terms of inputs used and that have gone through new technological paths. Technological advances, however, has not been uniform among firms. In fact, part of firms apparently seem to not move from their positions and get further away from the frontier. This section explores the hypothesis that firms have implemented different strategies to cope with increased competition. If it would be the case the observed increasing productivity dispersion could be related to the different modes of adaptation to the new economic conditions that firms followed. The underlying idea is that firms seek to adopt differet strategies to gain production flexibility to cope with external changes, drawing in different ways on labour market with different effect on performance (Michie and Sheehan, 2001, 2003; Arvanitis, 2005; Kleinknecht et al., 2006). Two main kinds of organizational flexibility firm can achieve. First, functional or internal flexibility, that is the ability to redeploy workers from one task to another, which require higher skills. Second, numerical or external flexibility, that is the ability to adjust the size of its workforce to fluctuations in demand by using workers who are not their regular, full-time employees. In fact, part of observed firms would have chosen 11 a cost-cutting strategy – a “low road” – putting, for instance, employees onto short-term contracts and/or part-time working, and accepting less skilled labour as solution to cope with the new competitive environment. In contrast, more dynamic firms would have chosen higher skilled labour to achieve a quality or value added advantage over competitors – they would go through the “high road”. Here, the interest is mainly in highliting some characteristics that may distinguish different “classes” of firms performance. To set off the analysis it seems usefull to create meaningful categories of performance. For this purpose, estimated productivity and efficiency so far illustrated, are used in a descriptive sense, that is as an indicators of a “residual” due to not observed factors. This residual can be explained both by esternal condition – industry affiliation, different access to external inputs due to location factors, etc. – and factors such as firms’ characterisitcs and different use of inputs – e.g., different factor costs if firms have access to input markets with different cost or quality. Then, firms are grouped, for each of the three periods, with respect to the industry average value of two variables: the level of efficiency at the beginning of each period and the dynamics of productivity observed in the same period. Figure 3 presents the obtained classifications. Productivity change t+!t Efficiency level t Low High Low High Laggards (1) Static Leader (2) N. obs: Period1 = 1439 Period2 = 1572 Period3 = 1693 Climbers (3) N. obs.: Period1 = 2468 Period2 = 2242 Period3 = 2090 N. obs.: Period1 = 2193 Period2 = 2303 Period3 = 2275 Dynamic leader (4) N. obs.: Period1 = 1247 Period2 = 1230 Period3 = 1289 Figure 3. Firms categories The dynamic leaders are firms closer to the technological frontier at the beginning of the period, which improve their productivity mainly through innovative strategies rather than efficiency improvements. The static leaders are firms close to the frontier, but with low productivity growth and therefore that tend, over time, to move away from the frontier. The climbers are firms with low initial efficiency, but that move rapidly towards the frontier and sometimes induce its shift. Productivity growth for these firms may be particularly fast as they can act on two factors: efficiency gains related to imitative processes and independent technological advancements. Finally the laggards are firms with low initial efficiency and low productivity growth, which nevertheless continue to be observed in the market throughout the period. A multinomial logit regression model is estimated in order to isolate some significant relatioships between a set of explanatory variales and the category: 12 " 4 % P ( y = j | x ) = exp ( x!k ) $1+ ! exp ( x! j )' $# j=2 '& (13) j = 2, 3, 4 rapresent the categories defined and x represents explanatory variables and controls. Obviously, for the reference category (1) we have: 4 " % P ( y = 1| x ) = 1! $1+ ! exp ( x! j )' $# j=2 '& (14) Two groups of variables are considered. The first group relates to the composition of the workforce and it contains a measure of unit labor cost (labour cost), the ratio between white and blue collars (skill ratio), and the share of part-time contracts on total employment (partime). The other variable is related to financial constraints (cash flow). We also considered the size – log of employees – and age – log of age – of firms. Table 2a,b shows mean and standard deviation and correlation matrix of the explanatory variables. Table 2a. Descriptive statistics Variable Labour cost skill_ratio partime cash flow size age Laggards Avg 20.7 0.46 0.043 393.0 54.5 22.5 Std 5.2 1.42 0.060 723.0 52.9 12.4 Static leader Avg 23.8 0.66 0.037 649.8 47.5 21.7 Std 7.4 2.53 0.053 1246.6 52.8 13.2 Climbers Avg 20.0 0.47 0.041 327.7 53.0 21.1 Dyn. leaders Std 5.6 1.28 0.058 576.6 49.6 12.4 Avg 23.8 0.88 0.039 564.3 44.0 20.9 Std 7.5 3.51 0.056 1151.8 46.6 12.7 Table 2b. Correlation matrix Correlation matrix Variable Labour cost skill_ratio partime cash flow size age Labour cost 1 0.167* -0.141* 0.225* 0.147* 0.256* skill_ratio partime cash flow size age 1 0.034* 0.078* -0.009 0.007 1 -0.075* -0.065* 0.038* 1 0.597* 0.092* 1 0.158 1 * significativity 5% Table 3 presents the estimated model using different sets of explanatory variables. The estimates refer to the log-odds ratio, i.e. the logarithm of the ratio of the probability of being in the category j over the probability of being in the baseline category. All models control for financial constraints, firm size, firms age, and include dummy variables to take into account periods, industry and the geographic area effects. Models differ 13 essentially for the introduction of control variables that allow refining the assumptions about the cost of labor. Model 1 takes into account only the unit cost of labor. It should be noted, however, that different labor cost may represent either different quality of labour employed, or, in the presence of a segmented labor market, may result from labour of the same quality but at a lower price. This hypothesis is consistent with the idea of twotier labor market (Boeri and Garibaldi, 2007). But, while in the former case the omission of an unobservable factor – the use of a single quality instead of two quality of labour force – can result in production inefficiency due to the use of labour of lower quality, in the latter case should imply only allocative inefficiency if firms do not adjust the composition of production factors, and then should affect only what part of the bestpractices frontier will be refence for each firm. In order to gain clearer interpretation of the results, Models 2 and 3 introduce, respectively, the ratio of white over blue collars that approximates the importance of upstream and downstream activities (Bugamelli et al., 2010), and the ratio of part-time contracts over total employees, which can be considered a measure of the use of flexible labor market (Arvanitis, 2005). Finally, in Model 4 introduces an interaction term between labor costs and times. As it can be seen in all models (Table 3), with reference to the probability of being a laggard, cash flow is positively associated with an increased probability of being leaders and reduces the probability of being a climber. Then the leaders will take advantage of less restrictive financial conditions, regardless of their productivity dynamics. However, the effect of cash flow, although statistically significant, is very low. The unit cost of labor acts in the same direction in all models: a higher cost of labor is associated to higher probability of being into the groups of firms close to the frontier (leaders) with respect to laggards. The effect is quite significant in terms of variation. The ratio of white over blue collars in Model 2 appears to act in the expected direction, and is particularly significant in explaining the probability of belonging to the group of dynamic leaders. The effect of this variable still persist in Model 3, after introducing the share of part-time contracts over the total number of employees. This latter variable reduces the probability of belonging into each of the categories with respect to laggards. The group of inefficient firms that do not increase productivity seem therefore to use more extensively labour flexibility. Finally, it should be pointed out that size significantly reduces the probability of being a leaders and reduces the probability of belonging to each of the different classes with respect to the laggards (but in terms of marginal effects, the effects of age are negative, once again, only for leaders). 14 Table 3. Multinomial logit estimates (log-odds ratios) . Reference category: Laggar Model 1 Variables labour cost skill ratio Static leaders Climbers Model 2 Dyn. leaders 0.11806*** -0.02078*** 0.12209*** (0.004) (0.004) (0.004) ! ! ! Static leaders Climbers Model 3 Dyn. leaders 0.11806*** -0.02283*** 0.11913*** (0.004) (0.004) (0.005) 0.03292* 0.02984 0.05720*** (0.018) (0.019) (0.018) ! ! ! partime ! ! ! period 2 * labour cost ! ! ! ! ! ! period 3 * labour cost ! ! ! ! ! ! cash flow age size Statistics Osservazioni Log-likelihood McFadden's Adj R2 Nagelkerke R2 LR !2 (dgr of freed.) estimates refer to log-odds ratio Standard errors in parenthesys *** p<0.01, ** p<0.05, * p<0.10 Dyn. leaders Static leaders Climbers Dyn. leaders 0.11672*** -0.02403*** 0.11783*** (0.004) (0.004) (0.005) 0.03586* 0.03254* 0.06005*** (0.018) (0.019) (0.018) -0.75662** -0.65536* -0.70499 (0.385) (0.353) (0.436) ! ! ! 0.12662*** (0.007) 0.03631* (0.019) -0.75431* (0.385) -0.01678* (0.009) -0.00976 (0.009) 0.00088*** (0.000) -0.37407*** (0.037) -0.96332*** (0.036) 0.00273 (0.007) 0.03374* (0.019) -0.71834** (0.353) -0.03858*** (0.009) -0.04252*** (0.009) -0.00030*** (0.000) -0.14653*** (0.036) 0.13546*** (0.033) 0.13823*** (0.007) 0.06071*** (0.018) -0.72638* (0.437) -0.03883*** (0.010) -0.02247** (0.010) 0.00084*** (0.000) -0.44027*** (0.042) -1.02293*** (0.040) ! Climbers ! ! 0.00088*** -0.00031*** 0.00085*** (0.000) (0.000) (0.000) -0.37521*** -0.14342*** -0.44413*** (0.037) (0.036) (0.041) -0.96826*** 0.13308*** -1.03350*** (0.035) (0.032) (0.039) 0.00088*** -0.00030*** 0.00084*** (0.000) (0.000) (0.000) -0.37386*** -0.13428*** -0.43571*** (0.037) (0.036) (0.042) -0.95791*** 0.13997*** -1.01723*** (0.035) (0.033) (0.040) 0.00088*** -0.00030*** 0.00084*** (0.000) (0.000) (0.000) -0.36909*** -0.12989*** -0.43112*** (0.037) (0.036) (0.042) -0.96107*** 0.13753*** -1.02017*** (0.036) (0.033) (0.040) yes yes yes yes yes yes yes yes yes dummies periods dummies industry dummies location constant Static leaders Model 4 2.12999*** (0.165) 0.87758*** (0.161) 21,258 -26569.454 0.076 0.207 4562.546 (66) 1.68603*** (0.186) 2.07361*** (0.167) 0.84948*** (0.163) 21,030 -26297.17 0.075 0.205 4474.089 (69) 1.64763*** (0.189) 2.13194*** (0.170) 0.89974*** (0.165) 21,030 -26294.682 0.075 0.205 4479.07 (72) yes yes yes 1.70268*** (0.192) 1.96952*** (0.197) 0.42065** (0.188) 21,030 -26274.224 0.075 0.207 4519.98 (78) 1.31008*** (0.221) To get a better idea of the impact of explanatory variables on the probability of belonging to a particular class, Table 4 presents, for Model 3 and 4, the marginal effects of variables on the probability of belonging to each category. Table 4. Marginal effects Model 3 Variables Model 4 Laggards Static leaders Climbers Dyn. leaders period 2 * labour cost -0.0109*** (0.0006) -0.00687** (0.0030) 0.122** (0.0542) ! 0.0209*** (0.0006) 0.00148 (0.0021) -0.0640 (0.0658) ! -0.0218*** (0.0007) 0.000424 (0.0024) -0.0310 (0.0616) ! 0.0118*** (0.0004) 0.00497*** (0.0012) -0.0265 (0.0514) ! period 3 * labour cost ! ! ! ! labour cost skill ratio partime cash flow age size dummies periods dummies industry dummies location -7.23e-05*** 0.000172*** -0.000187*** 8.75e-05*** (6.39e-06) (6.17e-06) (8.01e-06) (4.05e-06) 0.0502*** -0.0441*** 0.0289*** -0.0350*** (0.0054) (0.0059) (0.0058) (0.0046) 0.0956*** -0.164*** 0.169*** -0.101*** (0.0049) (0.0057) (0.0056) (0.0043) yes yes yes Laggards Static leaders -0.0141*** 0.0195*** (0.0010) (0.0010) -0.00700** 0.00142 (0.0030) (0.0021) 0.126** -0.0566 (0.0542) (0.0659) 0.00515*** 0.00204 (0.0013) (0.0013) 0.00430*** 0.00303** (0.0013) (0.00137 -7.26e-05*** 0.000172*** (6.38e-06) (6.18e-06) 0.0520*** -0.0431*** (0.0055) (0.0059) 0.0961*** -0.164*** (0.0049) (0.0057) Climbers Dyn. leaders -0.0181*** (0.0011) 0.000593 (0.0024) -0.0432 (0.0617) -0.00452*** (0.0014) -0.00682*** (0.0015) -0.000187*** (8.02e-06) 0.0264*** (0.0058) 0.169*** (0.0056) 0.0128*** (0.0007) 0.00498*** (0.0012) -0.0265 (0.0515) -0.00267*** (0.0010) -0.000511 (0.0010) 8.79e-05*** (4.06e-06) -0.0352*** (0.0046) -0.101*** (0.0043) yes yes yes It is now clearer that the skill ratio affects mainly the extreme categories, with a significant increase ot the probability of being a dynamic leaders and reduces the probability of being a laggard. This aspect seems particularly significant in light of the positive role that investment in upstream phases (e.g., design and product design) and downstream (e.g., marketing and sales) seems to have on firms competitiveness (Bugamelli et al. 2010). Moreover, it should be noted the sign of the interaction terms between labour cost and periods (Model 4). The marginal effect of changes in labour cost (Table 4), although negative in all periods, shows a different trend over time regarding the effect on the probability of belonging to categories 1 and 3. In the first case there is a trend towards a reduction of the negative impact while in the second case there is an amplification of the negative effect. This seems to suggest that after the phase of adaptation to the euro, climbers actually follow a pattern of reduction in unit labour cost with respect to leaders. In the end, it seems that laggard firms are on average characterized by higher age and adopt a defensive strategy, using labour of lesser quality and drawing from flexible labor market more extensively then other firms3. The climbers seem to use a mixture of strategies to reach the frontier, based on the servitization strategy and the use of less qualified labour. Leader firms, regardless of the dynamics of productivity, are younger, smaller, use more skilled labor and tend to be more tertiarized. 3 This evidence is coherent with Lucidi and Kleinknecht (2009), who found that Italian manufacturing !rms with a high share of "exible workers, lower costs of labour experienced signi!cantly lower rates of labour productivity growth from 2001 to 2003. The negative effect of firm size seems to be in constrast with a substantial proportion of the literature that shows a positive relationship between size and productivity. However, it is possible that in the periods examined, firms have experienced a process of downsizing. Moreover, we have seen how, in the first period, the increase in efficiency is also due to scale effects. As a consequence, it is likely that attempts to gain efficiency through adjusting the scale of production, in terms of number of employees, have hampered productivity improvement through the shift of the frontier. This result is indeed coherent with Hall et al. 2008. In fact, by analysing a panel of SMEs Italian manufacturing firm in the period 1995-2003, they found that larger and older firms were less productive. A final remark is necessary. Our exercise can only give some insights on the strategic nature of the observed productivity heterogeneity. Although the explanatory variables used are the level at the beginning of each period, concerns about endogeneity (cash flow, in the first place) still remain. The evidence put forward should be considered then only in a descriptive sense, nevertheless it suggests that the strategic nature of the observed heterogeneity has a foundation and that different groups of firms actually pursue differet strategy to adapt to new market conditions. 6. Conclusions Earlier studies on the Italian economic slowdown pointed to a generalized failure of the productive system to meet the challenges posed by the increasing globalization of markets. The analysis presented in this work suggests however that behind this generalized economic slowdown there could be high heterogeneity of firm behaviour. The evidences we presented are consistent with that obtained in related studies carried out with different methodologies. The approach we follow allows to isolate more precisely the component of productivity growth due to technological change. What emerges is evidence of a growing dualism. On the one hand some firms show sustained positive productivity dynamics, but at the same time, there is evidence that part of the system fails to keep pace with the group of innovators, with a consequent efficiency loss. In the latter part of the work it is questioned whether this dynamic may be related to different patterns of strategic adaptation. The evidence reported reinforce the hypothesis that firms follow different paths of adaptation and that different use of labour play decisive role in this process. Labour market reforms implemented in Italy in the ‘90s have indeed dramatically reduced the cost of use, and perhaps the quality, of newly hired workers. We hypothesized that part of the productive system has taken advantage of the emergence of a dualistic labour market. The availability of flexible labour, less expensive but also less skilled, has been for some categories of firms an easy solution to face competition. In contrast, more efficient and dynamic (dynamic leaders) firms have a higher unit cost of labour, are more outsourced, use less flexible labour, are younger and smaller. It is nevertheless difficult to assess the long run effectiveness of these different modes of adaptation. The first evidence we have, however, encourages a more careful analysis of this hypothesis. Finally, the nature of the analysis, based on a sample of firms along 11 years, can only observe firms that have survived and do not say anything about the effect of entry and exit on productiviy growth. In fact, it is known that firms turnover is a key 17 factor in explaining productivity growth (Bartelsman, Scarpetta and Hartiwanger, 2009): it would be necessary to move also in this direction to gauge better understanding on the origins of the long stagnation of productivity in Italy. References Aiello F., Pupo V., Ricotta F., (2009), Sulla dinamica della produttività totale dei fattori in Italia. Un'analisi settoriale. L'industria, 30, 3, 413-435. Arvanitis, S. (2005). Modes of labor flexibility at firm level: Are there any implications for performance and innovation? Evidence for the Swiss economy. Industrial and Corporate Change, 14, 993–1016. Banker, R.D., Charnes, A., Cooper, W.W., (1984). Some models for estimating technical and scale inefficiencies in data envelopment analysis. Management Science, 30, 1078-1092. Bartelsman E.J., Haltiwanger J.C., Scarpetta S. (2009), Cross-country differences in productivity: the role of allocation and selection. NBER working paper, n.15490 Boeri T., Garibaldi P., (2007), Two tier reforms of employment protection: A honeymoon effect?. The Economic Journal, 117, 521, 357-385. Brandolini A., Bugamelli M. (2009). Rapporto sulle tendenze nel sistema produttivo italiano. Questioni di economia e di finanza (Occasional Papers), n. 45, Banca d'Italia. Bugamelli, M., Schivardi, F. & Zizza, R. (2010), The euro and !rm restructuring, in A. Alesina & F. Giavazzi (Eds.), Europe and the euro, University of Chicago Press. Charnes, A., Cooper, W.W., Rhodes, E., (1978). Measuring the efficiency of decision making units. European Journal of Operational Research 2, 429-444. Coelli T.J., Rao D.S.P., O’Donnell C.J., Battese G.E., (2005), Introduction to Efficiency and Productivity Analysis, 2nd edition. Springer Daveri F., Jona-Lasino C., (2005), Italy's decline: getting the facts right”, Giornale degli economisti e annali di economia. 64, 4, 365-410. De Nardis, S. (2010), Imprese italiane nella competizione internazionale. Working paper ISAE, Roma. Dosi, G., Grazzi, M., Tomasi, C. and Zeli, A. (2011), Turbolence underneath the big calm? Exploring the micro-evidence behind the flat trend of manufacturing productivity in Italy. Small Business Economics.http://dx.doi.org/10.1007/s11187011-9326-7 Faini R., (2004), Fu vero declino? L'Italia negli anni '90. In V. Visco et al., Il declino economico dell'Italia: cause e rimedi. Milano: Bruno Mondadori. Färe R., Grosskopf S., (1996), Intertemporal Production Frontiers: With Dynamic DEA. Kluwer-Academic Publishers. Färe R., Grosskopf S., Norris M., Zhongyang Z., (1994), Productivity Growth, Technical Progress and Efficiency Change in Industrialised Countries. The American Economic Review, 84 , 1, 66-83. Hall B.H., Lotti F., Mairesse J., (2009), Innovation and productivity in SMEs. Empirical evidence for Italy. Small Business Economics, 33, 13-33. 18 Heshmati, A (2003). Productivity growth, efficiency and outsourcing in manufacturing and service industries. Journal of Economic Surveys, 7, 1, 79-112. Kleinknecht, A., R. M. Oostendorp, M. P. Pradhan and C. W. M. Naastepad (2006). Flexible labour, firm performance and the Dutch job creation miracle, International Review of Applied Economics, 20, 171–187. Kneip,A., Park, B., Simar, L. (1998), A note on the convergence of non-parametric DEA efficiency measure. Econometric Theory, 14, 783-793. Kneip, A., Simar, L., Wilson, P.W. (2008), Asymptotics and consistent bootstraps for DEA estimators in nn-parametric frontier models. Econometric Theory, 24, 16631697. Lucidi, F., Kleinknecht, A. (2009). Little innovation, many jobs: An econometric analysis of the Italian labour productivity crisis. Cambridge Journal of Economics, 34, 525-546. McDonald J., (2009), Using least squares and tobit in second stage DEA efficiency analysis. European Journal of Operational Research, 197, 792-798. Michie, J. and Sheehan, M. (2001), Labour market flexibility, human resource management and corporate performance. British Journal of Management, 12, 287 306. Michie, J. and Sheehan, M. (2003), Labour market deregulation, ’flexibility’ and innovation. Cambridge Journal of Economics, 27, 123–148. Milana C., Nascia L., Zeli A., (2008), Changes in multifactor productitvity in Italy from 1998 to 2004: evidence from firm-level data using DEA. EU KLEMS Working Paper Series, n. 33. Onida F. (2004), Se il piccolo non cresce. Le PMI italiane in affanno. Bologna: Il Mulino. Pianta M., Lucchese M., (2011), Crisis, cycles and innovation. Paper presented at the Conference Crises, Institutions and Labour Market Performance: Comparing Evidence and Policies, University of Perugia, Italy, 10 November 2011. Pianta, M., Vaona, A. (2007), Innovation and Productivity in European Industries, Economics of Innovation and New Technology, 16, 7, 485-99. Rossi, S. (2006), La regina e il cavallo. Quattro mosse contro il declino. Bari: La Terza. Sampaio de Sousa M.C. e B. Stosic, (2005), Technical efficiency of the Brazilian municipalities: correcting nonparametric frontier measurement for outliers. Journal of Productivity Analysis, 24, 157-181. Shephard, R. W. (1970), Theory of Cost and Production Functions. Princeton: Princeton University Press. Simar L. e Wilson P.W., (2008), Statistical inference in nonparametric frontier models: recent development and perspectives, in Fried H., Lovell C.A.K. e Schmidt S., (Eds.), The Measurement of Productive Efficiency, 2nd edition. London: Oxford University Press. Tulkens, H. and P. Vanden Eeckaut (1995), Non-parametric Efficiency, Progress and Regress Measures for Panel Data: Methodological Aspects. European Journal of Operational Research, 80, 474 - 499. Wheelock D.C., Wilson P.W. (1999). Inefficiency and productivity change in U.S. banking. 1984-1993. Journal of Money, Credit and Banking, 31, 2, 212-234. Zelenyuk, V. (2006), Aggregation of Malmquist productivity indexes. European Journal of Operational Research, 174, 1076–1086 19 APPENDIX A Table A1. Number of Firms and Employment for industries. Year, 2001 Firms a Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equip. Electronics Tansportation equipment Other manufacturing Manufacturing a) ISTAT Number % 8,328 7.2 13,929 12.0 4,869 4.2 3,281 2.8 9,838 8.5 352 0.3 3,797 3.3 5,993 5.2 6,399 5.5 20,545 17.7 15,879 13.7 11,291 9.7 2,697 2.3 8,716 7.5 115,914 100.0 Employees Ns. Database Number % 564 7.3 911 11.8 365 4.7 204 2.6 479 6.2 22 0.3 309 4.0 492 6.3 433 5.6 1,445 18.7 1,137 14.7 574 7.4 161 2.1 616 7.9 7.712 100.0 ISTAT Number 220,922 352,291 113,573 56,284 178,708 24,192 197,340 175,330 175,035 503,712 498,507 344,198 253,778 174,104 3,267,974 a % 6.8 10.8 3.5 1.7 5.5 0.7 6.0 5.4 5.4 15.4 15.3 10.5 7.8 5.3 100.0 Ns. Database Number % 25,404 6.2 51,645 12.6 19,971 4.8 9,071 2.2 21,419 5.2 1,045 0.2 17,313 4.2 26,858 6.5 21,676 5.3 77,814 19.0 62,991 15.3 31,104 7.6 10,691 2.6 32,288 7.8 409,290 100.0 Values refer to entire population Table A2. Input and output variables. Year 2006 Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipment Electronics Tansportation equipment Other manufacturing Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Turnover (Th !) Labour Services (mgl !) Row Mat. (Th !) 15,669.5 25,117.2 9,109.3 13,017.2 10,510.2 13,650.8 8,185.0 9,979.3 8,662.5 9,640.1 26,414.5 42,909.3 18,540.3 38,739.3 10,846.8 16,665.0 10,014.8 14,122.2 13,774.9 55,741.7 11,097.2 16,495.7 9,987.7 16,243.5 14,649.7 28,463.5 8,868.7 11,117.8 44.31 55.56 50.33 65.79 45.70 48.80 43.38 40.97 43.94 40.79 51.24 70.51 59.50 75.74 54.77 71.06 49.65 59.21 54.73 63.06 56.81 66.52 52.98 52.34 63.88 70.80 49.52 49.03 2,652.3 4,747.6 2,941.1 4,488.2 2,647.6 3,554.8 1,749.4 2,824.6 2,155.0 2,611.1 2,561.3 3,856.4 4,234.7 12,159.0 2,055.2 2,981.7 2,429.3 4,190.4 2,292.0 4,576.7 2,245.3 3,246.3 1,936.2 2,102.5 2,836.3 4,474.4 2,141.6 3,332.0 10,263.4 18,072.1 3,931.2 6,346.0 5,899.3 8,682.2 4,425.8 5,586.4 3,979.7 5,698.8 18,993.9 37,197.5 9,706.4 18,641.5 5,898.8 10,270.9 4,704.8 6,737.6 8,215.4 50,413.3 5,601.1 9,926.1 5,138.7 12,652.7 8,344.9 21,173.8 4,647.5 6,582.7 Tangible fix. assets (Th !) 3,490.4 5,209.3 1,629.7 3,459.8 1,186.6 1,738.5 2,030.9 3,055.5 2,153.4 3,512.9 8,176.1 25,167.8 2,764.4 4,774.9 2,484.2 7,319.7 2,459.3 3,500.9 2,330.1 4,813.5 1,556.4 2,514.6 1,258.7 2,210.3 2,115.5 2,719.7 1,752.0 2,960.9 Table A3. Malmquist index and its decomposition into Efficiency Change (EffCh) and Technological change (TeCh). Period 1996-2006. 1996-2006 Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipment Electronics Tansportation equipment Other manufacturing Malm EffCh TeCh 1.069 0.353 -0.004 0.268 0.090 -1.004 -0.032 0.490 0.068 0.273 1.019 1.224 0.558 0.196 0.050 -0.221 -0.171 0.226 0.534 -0.685 0.147 0.194 0.041 -0.033 0.143 -0.120 0.169 -0.175 1.018 0.574 0.167 0.043 -0.442 -0.326 -0.180 0.296 0.030 0.307 0.873 1.346 0.387 0.372 Table A4. Malmquist index and its decomposition into Efficiency Change (EffCh) and Technological change (TeCh). Subperiod: 1996-2000, 2000-2003, 2003-2006 1996-2000 2000-2003 2003-2006 Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipm. Electronics Tansportation equipment Other manufacturing Malm EffCh TeCh Malm EffCh TeCh Malm EffCh TeCh -0.849 -0.355 -1.001 -0.068 -1.579 -4.327 -1.223 0.041 -0.702 -0.218 -0.510 0.500 -1.010 -0.628 0.346 0.046 -0.106 0.636 1.228 -0.178 -0.165 0.887 0.677 0.654 0.636 0.703 0.197 0.201 -1.189 -0.400 -0.895 -0.701 -2.775 -4.155 -1.058 -0.835 -1.357 -0.864 -1.134 -0.200 -1.206 -0.826 1.624 0.823 0.860 0.975 1.908 8.473 1.794 1.738 0.391 1.239 0.767 1.200 0.523 0.055 0.311 -0.235 0.064 -0.641 0.876 -0.184 1.202 0.597 0.496 0.396 -0.058 0.258 0.768 -0.379 1.315 1.063 0.769 1.628 1.021 8.668 0.579 1.135 -0.104 0.841 0.830 0.948 -0.237 0.443 6.992 1.007 1.295 0.479 0.765 -4.522 -0.164 0.050 1.144 0.346 3.544 2.652 2.631 1.649 -0.511 -0.400 -0.401 0.596 -0.584 -1.587 -0.358 -1.078 -0.860 -1.305 -0.188 -1.622 -0.569 -0.349 7.509 1.408 1.676 -0.118 1.369 -2.959 0.194 1.143 2.033 1.667 3.740 4.337 3.209 2.005 Table A5. Efficiency Change (EffCh) and its decomposition into Pure Efficiency Change (PEffCh) and Scale Efficiency Change (SEffCh). Subperiods: 1996-2000, 2000-2003, 2003-2006 1996-2000 2000-2003 2003-2006 Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipm. Electronics Tansportation equipment Other manufacturing EffCh PEffCh SEffCh EffCh PEffCh SEffCh EffCh PEffCh SEffCh 0.346 0.046 -0.106 0.636 1.228 -0.178 -0.165 0.887 0.677 0.654 0.636 0.703 0.197 0.201 0.520 -0.215 -0.095 0.280 0.841 -0.069 -0.202 0.591 0.545 0.011 0.118 0.354 0.124 0.071 -0.170 0.268 -0.009 0.357 0.390 -0.108 0.040 0.303 0.133 0.646 0.521 0.361 0.084 0.134 0.311 -0.235 0.064 -0.641 0.876 -0.184 1.202 0.597 0.496 0.396 -0.058 0.258 0.768 -0.379 0.216 -0.129 0.122 -0.314 0.946 0.202 0.218 0.168 -0.021 0.186 0.299 -0.578 0.538 0.316 0.100 -0.100 -0.052 -0.323 -0.057 -0.385 0.991 0.432 0.522 0.216 -0.352 0.850 0.235 -0.686 -0.511 -0.400 -0.401 0.596 -0.584 -1.587 -0.358 -1.078 -0.860 -1.305 -0.188 -1.622 -0.569 -0.349 -0.412 -0.198 -0.486 0.336 -0.707 -0.208 0.194 -0.514 -0.857 -0.451 -1.025 -1.046 -0.150 -0.195 Tabla A6. Number of employee for different years Year Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipment Electronics Tansportation equipment Other manufacturing Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. Avg St.dev. 1996 2000 2003 2006 40.03 45.30 53.62 64.86 52.45 59.27 39.88 34.10 41.16 39.03 43.27 66.53 49.95 74.14 46.12 50.17 44.46 48.68 46.26 48.68 49.30 51.75 46.35 42.92 60.47 70.66 45.67 38.33 44.87 53.40 57.05 67.58 54.74 61.22 45.10 41.07 44.87 40.35 44.09 60.07 56.52 80.73 54.81 62.59 49.43 54.80 53.83 56.34 55.37 56.71 54.10 51.34 67.21 74.55 52.67 44.25 46.35 56.22 53.91 64.73 50.37 53.97 43.57 39.80 44.03 39.18 47.14 60.57 56.40 76.81 54.17 65.02 49.79 57.30 53.42 58.58 55.08 59.94 52.73 52.17 63.57 67.01 51.23 46.28 44.31 55.56 50.33 65.79 45.70 48.80 43.38 40.97 43.94 40.79 51.24 70.51 59.50 75.74 54.77 71.06 49.65 59.21 54.73 63.06 56.81 66.52 52.98 52.34 63.88 70.80 49.52 49.03 -0.096 -0.189 0.088 0.264 0.132 -1.382 -0.549 -0.562 0.000 -0.854 0.865 -0.580 -0.421 -0.149 Table A7. Technological Change (TeCh) and its decomposition into Pure Technological Change (PTeCh) and Scale Technological Change (STeCh). Subperiod: 1996-2000, 2000-2003, 2003-2006 1996-2000 2000-2003 2003-2006 Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipm. Electronics Transportation equipm. Other manufacturing TeCh PTeCh STeCh TeCh PTeCh STeCh TeCh PTeCh STeCh -1.189 -0.400 -0.895 -0.701 -2.775 -4.155 -1.058 -0.835 -1.357 -0.864 -1.134 -0.200 -1.206 -0.826 -1.042 0.171 -0.997 0.108 -1.755 -5.022 -0.134 0.174 -0.735 0.011 -0.415 0.537 0.138 -0.413 -0.096 -0.492 0.144 -0.742 -0.814 1.037 -0.638 -0.924 -0.530 -0.850 -0.680 -0.681 -1.096 -0.357 1.31 1.06 0.77 1.63 1.02 8.67 0.58 1.13 -0.10 0.84 0.83 0.95 -0.24 0.44 1.865 0.557 -0.746 1.004 0.789 8.555 1.631 1.420 0.353 1.052 0.552 2.015 -0.017 -0.161 -0.490 0.516 0.740 0.511 0.172 -0.031 -1.018 -0.262 -0.432 -0.252 0.284 -1.119 -0.214 0.614 7.509 1.408 1.676 -0.118 1.369 -2.959 0.194 1.143 2.033 1.667 3.740 4.337 3.209 2.005 1.604 1.642 0.873 2.142 -4.046 -0.350 1.202 2.716 1.152 5.095 4.465 3.222 2.018 -0.181 -0.004 -0.958 -0.727 1.120 0.610 -0.005 -0.650 0.560 -1.308 -0.027 -0.054 -0.050 Table A8. Technological Change (TeCh) and its decomposition into MaTeCh and Input Biased Technological Change (IbTeCh). Subperiod: 1996-2000, 2000-2003, 2003-2006 1996-2000 2000-2003 2003-2006 Industry Food and beverages Textile and wearing Leather Wood Paper and printing Petroleum Chemical Rubber and plastic mat. Non-met. mineral prod. Fabricated metal prod. Machinery and equipm. Electronics Transportation equipm. Other manufacturing TeCh MaTeCh IbTeCh TeCh MaTeCh IbTeCh TeCh MaTeCh IbTeCh -1.189 -0.400 -0.895 -0.701 -2.775 -4.155 -1.058 -0.835 -1.357 -0.864 -1.134 -0.200 -1.206 -0.826 -1.558 -0.645 -1.213 -1.133 -3.061 -11.724 -1.842 -1.071 -1.790 -1.315 -1.343 -0.505 -1.830 -1.047 0.406 0.270 0.348 0.461 0.329 9.503 0.958 0.252 0.458 0.479 0.218 0.318 0.742 0.236 1.315 1.063 0.769 1.628 1.021 8.668 0.579 1.135 -0.104 0.841 0.830 0.948 -0.237 0.443 0.380 0.897 -0.252 1.004 0.689 4.349 0.311 0.918 -0.380 0.589 0.661 0.729 -0.850 0.188 0.958 0.183 1.045 0.641 0.353 4.129 0.279 0.237 0.283 0.265 0.171 0.220 0.655 0.275 7.509 1.408 1.676 -0.118 1.369 -2.959 0.194 1.143 2.033 1.667 3.740 4.337 3.209 2.005 -0.594 1.148 1.236 -0.657 1.071 -7.801 -0.231 0.918 1.774 0.290 3.396 3.459 2.414 1.593 9.473 0.266 0.780 0.577 0.353 5.914 0.449 0.236 0.266 1.421 0.347 0.870 0.789 0.416 APPENDIX B: Outliers detection strategy We use a method based on the concept of leverage, that is the effect produced on the efficiencies of all the other firms when the observed firm is removed from the dataset (Stosic and Sampaio de Souza, 2005). The underlying idea is that outliers are expected to display leverage much above the mean leverage. The leverage measure is then calculated for each firm, and it is used to detect and to automatically eliminate outliers from the dataset. Formally, the leverage is measured as the standard deviation of the efficiency estimates relative to the full sample without the evaluated observation, and the inefficiency estimates relative to the full sample. I order to calculate it, the most straightforward possibility is to use jackknife resampling. One first applies DEA to each of the firms using the unaltered, original dataset to obtain the set of efficiencies {"k , k = 1,...,K} . Then a “leave-one-out” strategy is employed: one by one, each firm is successively removed from the dataset, and each time the set of efficiencies {"ki* , k = 1,...,K, k # i} is recalculated, where i=1,…,K indexes the removed firm. ! Finally, the leverage of the i-th firm may then be defined as: % (" ! (17) leveragei = k= S L ,k$i * ki # "k ) K #1 A higher leverage value provides evidence for an influential observation. We perform the outlier procedure with respect to eahc year forntier at industry level. This ! approach however is extremely computationally intensive. Therefore, as in Stosic and Sampaio de Sousa (2005), we use a more efficient stochastic procedure, which combines bootstrap resampling with the above jackknife strategy. In our procedure we reduce the computational burden of the procedure proposed by Stosic and Sampaio de Sousa by calculating the leverage only for those firms that are efficient given the extracted subsample, while for inefficient firm the leverage is set to zero. This because the elimination of an inefficient firm from the data set under analysis has no effect on the efficiency of all other firms. The procedure is as follows1: [1] Loop the following steps ([1.1]–[1.5]) B times and obtain estimates of the B leverage A = {lbi }b=1 "i # SLEff [1.1] extract randomly without reemission a subset SL of cardinality L from the original data set S of cardinality K [1.2] calculate the efficiency " k #k $ SL ! [1.3] partition SL = SLEff " SLIneff [1.4] loop the following step ([1.3.1]) "j # SLEff [1.4.1] remove firm the current firm j " SLEff and calculate the ! efficiency {" ki* : k = 1,...,LIneff } where i is the removed firm. ! [1.4] calculate the leverage ! ! ! 1 The procedure has been implemented using the statistical package R. ( ** li = ) * *+0 % (" k= S L ,k$i * ki # "k ) K #1 &i ' SLEff &i ' SLIneff B [2] calculate the average leverage li = " lbi n i , where ni is the number of time b=1 ! firm i is extracted. K [3] calculate the global average leverage: l = " li K i=1 ! The procedure requires the choice of the cardinality of the subsample extracted in each iteration and the number of repetitions. Stosic and Sampaio suggest a cardinality that is ! dataset S and a number of repetitions that is 10-20% of the cardinality of the entire greater or equal to 1000. In this paper we chose L=0,2 and B=1000. In order to take into account the number of observations used, we define as threshold levarage the quantity l log(K ) . Observation with average leverage greater than the threshold was omitted. ! APPENDIX C: Capital stock estimation procedure To obtain estimates of capital stock the procedure uses historical values of tangible assets and adapts the perpetual inventory method. Specifically, the procedure assumes that assets are purchased and then replaced over a certain period of time depending on their estimated average duration. The first step consists in estimating capital for the base year – the firs year. It is assumed that assets in first year has been bought gradually over the past and then replaced in subsequent years. For the base year, historical value of tangible assets is divided for the estimated average duration and each portion is deflated using the specific deflator for the related period. The estimated capital stock for the first year (base) is as follow: base $k kbase 1 kbase 1 kbase 1 1 ' base k̂base = ! + ! +!+ ! =& ! # ) d deflbase d deflbase"1 d deflbase"d % d i=base"d defli ( where kbase is the historical value of tangible assets for base year, d it the average assets duration calcuated at 2-digit level and defli is the deflator for year i. Deflators are obtained by processing ISTAT data. In particular, given the unavailability of the investments series at the 2-digit level, the deflator are common to all firms and are built as the ratio of the monetary value of total investments at current prices, in a given year, over the corresponding value in the chained series, and the base year is 2000. For subsequent years, capital stock is splitted into two part: the estimated quantity of capital that still survive at time t and the estimated new investment at time t. Thus for year t the adjusted value of tangible assets kt is as follow: " " k % ( k %+ kt = $ kt!1 ! t!1 ' + *kt ! $ kt!1 ! t!1 'dt!1 & ) dt!1 &, # # where kt!1 is the historical value of tangible assets at time t-1 and dt!1 is the average duration caculated for year t-1. Specifically first component of the right hand of equation above represents the estimated volume of tangible assets that still survive at time t, while the second component the estimated new investment at time t. Therefore, the estimated stock of capital k̂t a time t is obtained as follows: "" % " " k % k %% $ $ kt!1 ! t!1 ' t!1 ' $ kt ! $ kt!1 ! t!1 ' ' dt!1 & dt!1 & ' 1 ' $ # # k̂t = $ () + $ (d !1) ' $ ' deflt i=t!d defli $ ' $ ' # & # &