No 2015-19 – October Working Paper Should everybody be in services? The effect of servitization on manufacturing firm performance Matthieu Crozet & Emmanuel Milet Highlights In 2007, about 70% of the French manufacturing firms produced some services for third parties. This share is growing over time. These servitized firms are larger (in terms of total production and employment), produce more goods and are more profitable. There is a causal impact of servitization on firm performance: Firms that start selling services experience an increase in their profitability, employment, total sales and sales of goods. These positive effects of servitization strategies are mainly visible for small businesses. CEPII Working Paper Should everybody be in service? Abstract The servitization of the manufacturing sector refers to the evolution of manufacturers' capabilities to offer services as a complement to or a substitute for the goods that they produce. A vast literature has described these strategies and has shown that this phenomenon is widespread and growing in most developed economies. However, very little systematic evidence of the extent or consequences of servitization based on a comprehensive dataset of firms exists. In this paper, we provide such evidence using exhaustive data for French manufacturing firms between 1997 and 2007. We find that the vast majority of French manufacturers sell services in addition to producing goods. The shift toward services is growing steadily but at a slow pace. We also provide evidence of a causal impact of servitization on firm performance. Controlling for various sources of endogeneity bias, we find that firms that start selling services experience an increase in their profitability between 3.7% and 5.3%, increase their employment by 30%, increase their total sales by 3.7%, and increase their sales of goods by 3.6%. The results hold for most industries, although some heterogeneity exists. Keywords Servitization,Deindustrialisation,Firm performance. JEL L23, L25, L6. Working Paper CEPII (Centre d’Etudes Prospectives et d’Informations Internationales) is a French institute dedicated to producing independent, policyoriented economic research helpful to understand the international economic environment and challenges in the areas of trade policy, competitiveness, macroeconomics, international finance and growth. CEPII Working Paper Contributing to research in international economics © CEPII, PARIS, 2015 All rights reserved. Opinions expressed in this publication are those of the author(s) alone. Editorial Director: Sébastien Jean Production: Laure Boivin No ISSN: 1293-2574 CEPII 113, rue de Grenelle 75007 Paris +33 1 53 68 55 00 www.cepii.fr Press contact: presse@cepii.fr CEPII Working Paper Should everybody be in services? Should everybody be in services? The eect of servitization on manufacturing rm performance Matthieu Crozet  and Emmanuel Milet y There are no such thing as service industries. There are only industries whose service components are greater or less than those of other industries. Everybody is in services. Theodore Levitt (1972). 1. Introduction The servitization of the manufacturing sector refers to the evolution of manufacturers capabilities to oer services, as a complement to or a substitute for the goods they produce. This trend is not recent and has been identied and documented since the 1980s (Vandermerwe and Rada, 1988). It is observed in all OECD countries and also in developing countries (Pilat et al., 2006; Neely et al., 2011). Examples of manufacturing rms selling services are numerous: from small businesses oering repair and after-sales services to Rolls Royce, which made power by the hour  a package of support services 1 or Apple, whose strategy is to oer 2 to consumers an ecosystem combining physical devices with online services. for aircraft engines  a core element of its strategy, The deeper integration of the production of goods and services is highly relevant for policymakers in high-income countries who worry about the decline of manufacturing production and employment in their economies. 3 Economic analyses based on a representation of the economy as a collection of independent sectors often view the decline of the manufacturing sector as an ineluctable shift of resources toward the services sector (Baumol, 1967; Ngai and Pissarides, 2007; Acemoglu and Guerrieri, 2008). However, this representation as well as the ensuing industrial policies neglect the fact that the boundary between manufacturing and services is very blurry 4 and that complementarity between services and manufacturing may be key to economic success.  Univ Paris Sud, Universite Paris-Saclay and CEPI, (matthieu.crozet@cepii.fr) y University of Geneva, (emmanuel.milet@unige.ch) 1 Rolls-Royce earns its keep not just by making world-class engines, but by selling power by the hour"  a complex of services and manufacturing that keeps its customers' engines burning. If it did not sell services, Rolls-Royce could not earn enough money from selling engines", The Economist (Jan. 8th, 2009). 2 Between 2002 and 2010, Apple sold over 206 million iPods and over one billion songs through the iTunes Music Store (Benedettini et al., 2010). 3A recent report by the European Commission argues that European Commission (2014) A digital tran- sition is underway across the global economy and industrial policy needs to integrate new technological opportunities such as cloud computing, big data and data value chain developments, new industrial applications of internet, smart factories, robotics, 3-D printing and design. 4A fact underlined decades ago by Stigler (1956): There exists no authoritative consensus on either the boundaries or the classication of the service industries. increasing importance of the services sector. 1 See Fuchs (1968) for an early discussion of the CEPII Working Paper Should everybody be in services? Of course, the shift toward services has important consequences for rms; it aects their business models and how they approach consumers (Oliva and Kallenberg, 2003; Reinartz and Ulaga, 2008; Cusumano et al., 2015). This shift also provides a way to restore 5 Wise and Baumgartner manufacturers' competitiveness in both local and global markets. (1999) argue that Downstream [service] markets oer important benets besides large new sources of revenue. They tend to have higher margins and to require fewer assets than product manufacturing. And because they tend to provide steady service-revenue streams, they're often countercyclical. Clearly, in manufacturing today, the real money lies downstream, not in the production function. Previous studies of the consequences of servitization have identied various channels through which rms can benet from this strategy. Servitization can enable rms to dierentiate their product from those of their competitors (Baines et al., 2009), increase customer loyalty (Baines and Lightfoot, 2013), and lead to higher market values (Fang et al., 2008) or higher protability (Neely, 2008; Suarez et al., 2013; Visnjic et al., 2014). However, little systematic and robust evidence of the impact of servitization on rm performance exists, and this question remains controversial. Prior research has shown that most of the expected benets of servitization (higher revenues, higher protability) does not materialize in many cases. 6 Furthermore, most of the available empirical evidence is based on rm-level case studies or a limited sample of relatively large rms. These approaches have the advantage of allowing in-depth analysis of the business strategies and channels through which servitization operates. However, they lack the general validity that would allow inferences to be drawn from their results. Our paper is complementary to the existing literature in this regard. Our study is based on a comprehensive sample of rms, which covers all manufacturing sectors and naturally includes a large proportion of micro and small businesses. Our data contains detailed balance sheet information for more than 50,000 servitized and non-servitized French manufacturing rms over the 19972007 period. A key feature of our database is that it provides information on the sales (to third parties) of goods and services separately. This very large database allows the precise quantication of the evolution of the servitization of French manufacturing over the course of a decade; it also oers a means to estimate the causal impact of servitization on rm performance precisely, controlling for self-selection eects and other sources of endogeneity bias. 5 Cusumano et al. (2015) distinguish between services that are oered as complements to the product sold by the rm and those that are substitutes. Within complementary services, they further distinguish between smoothing services whose purpose is to ease the purchase of the product (nancing, insurance, basic training) and adapting services whose purpose is to alter the good to the specic needs of customers. While smoothing services can easily be standardized, adapting services are highly customized as the knowledge required to provide the service is dicult to separate from detailed knowledge of the product itself  . The Power by the Hour Rolls Royce product is an example of services that substitute for the purchase of the good. Customers buy the use of the engine rather than the engine itself, while Rolls Royce ensures that it is functional at any time. 6 This service paradox is described by (Gebauer et al., 2005) as follows: Most product manufacturers are confronted with the following phenomenon: Companies which invest heavily in extending their service business, increase their service oerings and incur higher costs, but this does not result in the expected correspondingly higher returns. Because of increasing costs and a lack of corresponding returns, the growth in service revenue fails to meet its intended objectives. We term this phenomenon the `service paradox' in manufacturing companies. Visnjic et al. (2014) argue that the service paradox is likely to arise when rms move from product-related services to customer-oriented services. 2 CEPII Working Paper Should everybody be in services? The empirical literature has typically found an ambiguous relationship between servitization and rm performance. Fang et al. (2008) use data on 477 publicly listed manufacturing companies and nd a U-shaped relationship between the share of services of total sales and rm market value. Benedettini et al. (2013) analyze the characteristics of about 200 manufacturing rms from the manufacturing sector. After controlling for rm age and size, they nd a negative correlation between the number of services oered by rms and their survival probability. Eggert et al. (2011) examine 414 German companies in the mechanical engineering industry and link product innovation and servitization to rm protability. They nd that when combined with product innovation, oering services supporting the product leads to higher protability. Finally, the work closest to ours in terms of data and methodology is the analysis provided by Suarez et al. (2013). They look at the eect of servitization on operating prots using a sample of slightly fewer than 400 rms in the prepackaged software products industry and nd a convex relationship between the share of services of total sales and overall operating margins. Their study covers a longer period than ours (19902006), but it is limited to one industry. We depart from several features of their analysis. First, we use a large sample of French rms that covers all manufacturing industries and allows us to assess how the impact of servitization varies by industry and type of rm. Second, we focus on the shift toward services rather than on the degree of servitization. Indeed, our results indicate that rm performance is mainly aected by the decision to engage in the provision of services rather than by their importance relative to total sales. Third, we look at several measures of performance: protability, employment, total sales, and production sales of goods. This latter indicator of performance leads to us to discuss the complementarity or substitutability of goods and services. Fourth, thanks to the detailed nature of our data, we can implement a very precise micro-econometric estimation strategy, which addresses self-selection bias and reverse causality. Finally, our contribution is twofold. First, we exploit our comprehensive database to document the extent of servitization in the French manufacturing sector. We show that in all French manufacturing industries, the share of services of total sales has increased substantially between 1997 and 2007. This increase is driven by two components: faster growth among servitized rms and a tendency for each rm to increase its share of services of total sales. Second, we assess the causal eect of engaging in services sales on rm performance. We explicitly tackle unobserved heterogeneity using a lagged dependent variable (LDV) model and reverse causality issues using instrumental variables. We nd that rms that start selling services experience a signicant boost in their protability, ranging between 3.7% and 5.3%. Their employment increases by 30%, total sales by 3.7%, and sales of goods by 3.6%. These results hold for most industries, although some heterogeneity exists. Additionally, these results are primarily driven by small and medium enterprises, while the estimation for a sample of large rms produces less signicant results. In the next section, we describe the dataset. We then describe the change in servitization in French manufacturing industries in section 3. In section 4, we present our empirical strategy and our results. We conclude in section 5. 3 CEPII Working Paper 2. Should everybody be in services? Data We use rm-level information from the BRN (Bénéce Réels Normaux) database. The database is compiled by the French scal authority (Direction Générale des Impôts) and provides rm information such as employment, value added, capital stock, prots, investments, industry classication, and geographic location. Of particular interest for this paper, the BRN dataset reports detailed information on rm sales. Individual sales are split into three mutually exclusive categories: sales of production of goods, sales of production of services and sales of merchandise (goods purchased and sold without transformation). Note that these are sales to third parties, i.e., to consumers outside of the rm. 7 It is important to note that in this paper, we are interested in the servitization of French manufacturing rms, i.e., in the fact that manufacturing rms sell services to third parties. We are not interested in the production of services for own accounts. 8 Our data cover the 19972007 period. The raw dataset provides information on 67,385 manufacturing rms. The average rm employs 55 workers and generates a turnover of e12 million. This dataset is very large, but individual data are noisy and sometimes report values that we consider highly dubious. For instance, some rms change their industry classication every year, moving from one 2-digit sector to another. This complicates the design of an appropriate control group, as we want to compare rms operating in the same industry. The dataset also contains information on rms that report no production, no value added, or no employment. To cope with these issues, we trim the data based on several factors. First, we keep rms reporting strictly positive sales, employment or value added. step reduces the size of the dataset quite substantially. This Second, we select those rms whose capital to labor ratio and value added per worker are not greater than a hundred times the median ratio in their industry. 9 We are left with 50,530 manufacturing rms. In this sample, the average rm employs 60 workers and has total sales of approximately e13 million. In our sample of manufacturing rms, 76% report selling some services. These rms account for approximately 90% of the total value added and employment in our sample. Among the rms that report positive sales of services, 22% report more sales of services than sales of (produced) goods, and 12% report only selling services. There are several explanations for these two somewhat surprising facts. Some rms may be misclassied and registered as manufacturing rms although their main activity is services. Other rms may have given up the production of goods to focus on the provision of services while still selling goods that they buy from other rms. 7 The It is important to note that French dataset does not provide information on the type of service sold by rms. Services and product sales are indistinctly exports or domestic sales. Note that for rms that belong to a group, they can be sales to other aliates or to subsidiaries of the group. 8 Lodefalk (2013) considers the purchase and the production of services in Swedish manufacturing rms over the 19752005 period. 9 Firms report such extreme values of these ratios for two reasons. The rst obvious reason is misreporting; the second is related to how rms manage their capital. Consider the following example: For tax purposes, a rm may decide to create an entity whose only purpose is to own its real estate assets. In this setting, the rst rm is producing goods and employing workers but appears to have little or no capital. The second rm, which is entirely linked to the rst, has a (potentially large) capital stock with few (if any) workers. Depending on how the boundaries of the rms are dened, we are left with two apparently distinct entities, both with capital to labor ratios that do not actually reect the activities of the rm. 4 CEPII Working Paper Should everybody be in services? rms are not automatically reclassied when their main activity changes partly because collective labor agreements are dened at the sector level, which can make reclassication costly and cumbersome for both employers and employees. To better describe this dual activity of French manufacturing rms, we dene the service intensity of a rm as the share of services of total production sales. The service intensity ranges from 0 (pure goods producers) to 1 (pure services producers). In gure 1, we present the kernel distribution of service intensity (on a log scale) for the manufacturing rms in our sample in 1997 and 2007. The striking feature of this distribution is its bi-modality. This feature is present in 10 both years and is observed in each manufacturing industry (see gure 2). Most of manufacturing rms are mainly goods producers (they produce and sell more goods than services). In 2007, rms with a service intensity below 50% accounted for 84% of the rms in our sample and for 90% of both value added and employment. Figure 1 also reveals that the distribution remained quite stable between 1997 and 2007. Density (log scale) 2 4 6 810 Figure 1  Distribution of the Service Intensity (share of services of production sales) for French Manufacturing Firms in 2007 0 .2 .4 .6 Share of services in total output 1997 .8 1 2007 To dig deeper into the changes in the distribution of service intensity over time, we computed a transition matrix between 1997 and 2007 (see table 8 in the appendix). In this table, we retained a constant sample of rms and allocated them to bins based on their initial service intensity in 1997 and their service intensity in 2007. To understand how to interpret the gures reported in the table, let us consider the rst row of the table: We nd the share of rms that had a service intensity of exactly 0% in 1997. Adding all shares reported in this row, the table indicates that this was the case for 22% of the rms in our sample. By 2007, a majority of these rms (11.98% of the total sample) remained fully 10 Our database covers 21 2-digit industries. In the econometric analyses presented in sections 3 and 4, we systematically control for 2-digit industry xed eects. However, to facilitate the exposition, we group the 2-digit industries into 7 broad categories. 5 CEPII Working Paper Should everybody be in services? Density (log scale) 5 1015 Figure 2  Distribution of the Service Intensity by Industry in 2007 0 .2 .4 .6 Share of services in total output I V II VI .8 III VII 1 IV We grouped industries into large sectors using the NACE-Rev1 industry classication (indicated in parentheses). I: Manufacture of food products, beverages and tobacco (15, 16). II: Manufacture of textiles and leather products (17, 18, 19). III: Manufacture of wood and wood products; manufacture of pulp, paper and paper products; publishing and printing (20, 21, 22). IV: Manufacture of chemicals, chemical prod- ucts and man-made bers; manufacture of rubber and plastic products (24, 25). V: Manufacture of other non-metallic mineral products, basic metals and fabricated metal products (26, 27). VI: Manufacture of machinery, electrical, optical and transport equipment (29, 30, 31, 31, 33, 34, 35). VII: Manufacturing, n.e.c. We omitted rms in the manufacture of coke, rened petroleum products and nuclear fuel industry (23), as only 4 rms existed in 2007. specialized in the production of goods, while the rest (about 10% of the total sample) were selling some services in 2007. Among the latter, a vast majority (7.42% of the sample) had sales of services that accounted for less than 10% of their total production sales. This is a salient feature of the matrix: Most rms do not change their production mix much. Approximately 60% of rms lie on the diagonal of the table. Only 23% of rms are located strictly above the diagonal, meaning that they substantially increased their service intensity between 1997 and 2007, and only 17% of the rms decreased their service intensity. Finally, very few rms completely changed their production mix during this period: Only 3.5% of the rms moved from a low service intensity in 1997 (below 10%) to more than 90% of services in 2007 (0.25+1.30+0.42+1.57=3.54). Together, gures 1 and 2 and table 8 convey the following message. two distinct types of rms in the French manufacturing sector: We encounter Firms that are mainly goods producers (with a service intensity below 50%) and those that are specialized in the provision of services. The distribution of these two types of rms is quite stable over time, and very few rms move from one type to another. As the focus of this paper is the production of services by manufacturers, we consider rms that are mainly good producers (i.e., those that have always a service intensity below 50%). All rms that sell 6 CEPII Working Paper Should everybody be in services? more services than goods in at least one year are excluded from our sample. This leaves us with a sample of 39,814 manufacturing rms, which remains quite similar to the raw dataset. In this sample of goods producers, the average rm employs 66 workers and generates e14 million in total sales. The average rm in this sample is slightly larger than in the raw dataset, although the dierence is not statistically dierent. It is important to notice that excluding rms with a share of services above 50% has no consequences for our econometric analysis of the impact of servitization presented in section 4. As our identication strategy relies on rms switching from zero to positive sales of services, rms with a service intensity above 50% typically provide services every single year in our sample and thus do not contribute to the identication process. 3. The servitization of French manufacturers This section gives an overview of the degree of servitization of French manufacturing over the decade from 1997 to 2007. Table 1  Summary Statistics All rms (1) # Firms Servitized rms 1997 2007
97-07 1997 2007
97-07 25,660 22,675 -1.23 17,826 15,740 -1.24 69.4 69.4 1,443 1,274 86.9 89.9 Share (%) (2) Employment Total (thousand) 1,661 1,417 -1.58 Share (%) Average (3) Turnover e, Total ( million) 64.7 62.5 -0.35 80.9 81.0 +0.0 294.3 378.0 +2.53 261.0 350.2 +2.99 88.7 92.6 Share (%) Average ( (4) e, thousand) Production of goods e, Total ( million) 11.5 16.7 +3.81 14.6 22.3 +4.27 261.4 325.2 +2.21 229.0 298.0 +2.67 87.6 91.6 Share (%) Average ( (5) (6) e, 10.2 14.3 +3.48 12.8 18.9 +3.95 Average (%) 47.44 49.0 +0.31 48.4 50.5 +0.42 Average (%) 3.1 3.2 +0.33 4.5 4.6 +0.34 Median (%) 0.5 0.6 +2.47 1.4 1.6 +1.76 0.07 0.07 0.08 0.08 Protability thousand) -1.23 Service intensity Std. dev.
97-07 corresponds to the annualized growth rate between 1997 and 2007. The sample is rms producing mainly goods (i.e., whose service intensity is below 50% over the period). Servitized rms report strictly positive sales of services. In table 1, we present some descriptive statistics for the population of servitized and 11 In the left part of the table, non-servitized French manufacturing rms in our sample. 11 All these gures are computed from the sample of rms that are mostly producers of goods (i.e., those 7 CEPII Working Paper Should everybody be in services? we show statistics for servitized and non-servitized rms, and we restrict the sample to servitized producers in the right part of the table. The rst line illustrates the rapid deindustrialization of French economy. Between 1997 and 2007, the number of rms in our sample decreased by almost 12% (equivalent to an annual growth rate of -1.23%), and the number of workers employed in our sample of manufacturing rms decreased by 14.7% (-1.58% per year). Table 1 also shows that servitized rms are, on average, larger than pure manufacturers. In 2007, they employed 81 workers, on average, compared to 62.5 for pure manufacturers. Servitized rms are also larger in terms of turnover, they produce and sell more goods, and they are more protable. All these dierences will be studied in detail in the next section. Figure 3 better illustrates the extent of servitization across manufacturing industries, using three dierent indicators. Panel (a) shows the proportion of manufacturing rms that produce services in 1997 and in 2007. It conrms that servitization is a quite common strategy among French manufacturing rms: Almost 70% of the rms in our sample produce some services for third parties. This gure varies substantially by sector, ranging from 54% in the food, beverage and tobacco industry to 88% in the chemical and plastic products industry. The share of servitized rms has increased in every industry between 1997 and 2007, with the exception of the wood, paper and printing industry. While a very large majority of rms are servitized in all industries, most of them sell very few services. This pattern is visible in panel (b), which displays the average service intensity in each industry. In 2007, the service intensity is equal to 3.2% for the manufacturing sector as a whole. Again, there is some heterogeneity across manufacturing industries. For the average rm in the food, beverage and tobacco industry, services account for 1.3% of production sales, while they account for 5% in the mechanical and electrical equipment industry (which here includes optical and transport equipment). 12 Finally, panel (c) shows the importance of services to the total production of manufacturing industries. It reports the average service intensity weighted by the production of each rm. These gures are, on average, larger than those in panel (b), suggesting that larger rms have on average higher levels of service intensity. In 2007, services accounted for 5.5% of the total production sales of the manufacturing sector compared to 4.2% in 1997. Services accounted for around 2.5% of total production sales in the food, beverage and tobacco industry and up to 8.1% of production sales in the mechanical and electrical equipment industry in 2007. This measure of the scope of servitization shows a steady rise over the decade, as shown in gure 4. The gure displays the evolution of the weighted average service intensity between 1997 and 2007 along with the share of employment at servitized rms. We take 1997 as the reference year, so the vertical axis can be read as growth rates. The weighted average service intensity was 30% larger in 2007 than in 1997, and the share of employment at servitized rms grew by an average of 0.3% per year over the period. with a service intensity below 50%). Table 7 in the appendix displays exactly the same statistics computed from the complete sample of rms. 12 Logically, retaining the (relatively few) rms that produce more services than goods in the sample greatly changes the average level of service intensity but does not aect the median value much. For 2007, table 7 in the appendix reports an average service intensity of 23.1% and a median of 3.2%. 8 CEPII Working Paper Should everybody be in services? Figure 3  The Extent of Servitization in French Manufacturing Industries (a) Share of servitized rms (b) Average service intensity (unweighted) 1-Food, Beverage, Tobacco 1-Food, Beverage, Tobacco 2-Textile, Leather 2-Textile, Leather 3-Wood, paper, printing 3-Wood, paper, printing 4-Chemical, plastics 4-Chemical, plastics 5-Mineral, metal products 5-Mineral, metal products 6-Mechanical & electrical equip. 6-Mechanical & electrical equip. 7-n.e.c. 7-n.e.c. All manufacturing All manufacturing 0 .2 .4 .6 2007 .8 0 .01 .02 1997 2007 (c) Share of services in industry production 1-Food, Beverage, Tobacco 2-Textile, Leather 3-Wood, paper, printing 4-Chemical, plastics 5-Mineral, metal products 6-Mechanical & electrical equip. 7-n.e.c. All manufacturing 0 .02 .04 2007 .06 .08 1997 1 .99 1 1.1 1.01 1.2 1.02 1.3 1.03 1.4 1.04 Figure 4  Aggregate Servitization and Employment: 19972007 1997 1999 2001 2003 2005 Average servitization - Weighted (left axis) Employment in servitized firms (right axis) 2007 .03 .04 1997 .05 CEPII Working Paper Should everybody be in services? The global trend toward services shown in gure 4 contains two potential sources of change: a generalized shift toward services in individual rms and a composition eect due to the fact that rms with a high service intensity may grow faster than other rms. We isolate the rst source of variation by estimating the following equation: Service Intensity = +  +  ; it t (1) it t; , which controls for any rm characteristic constant over time, is a rm xed eect;  is a set of year dummies; and  is the error term. We omit the dummy for the year 1997 so that the estimated coecients on the dummies  measure the yearly average change in where Service Intensity i it is the share of services of total sales of rm i at time i t it t service intensity within each rm with respect to 1997. We display the results graphically in gure 5 where we plot the coecients on each  t along with a 95% condence interval. A positive (and statistically signicant) coecient means that, on average, rms have increased their service intensity with respect to their initial level in 1997. The dashed line represents an estimation with simple OLS, while the solid line shows the estimates obtained from a linear regression wherein observations have been weighted by the average employment of the rm over the period. All the coecients are positive and statistically signicant and increase over time. This means that the growing importance of services in the total production of manufacturing rms is not entirely driven by faster relative growth among servitized rms. On the contrary, each rm has increased its service intensity between 1997 and 2007 by an average of 1.5 percentage points. In 1997, the (weighted) average service intensity was 4.2%. An average increase of 1.5 percentage points in each rm represents an average increase of 35% over the decade (or 3% per year). 0 .005 .01 .015 .02 Figure 5  Firm-level Servitization: 19972007 1998 2001 2004 Weighted estimates 4. 2007 Non-Weighted estimates The impact of servitization on manufacturing rm performance In this section, we analyze the interaction between servitization and rm performance. The following two subsections address two questions: Do servitized rms outperform pure manufacturers, and do rms that shift toward services improve their performance? 10 CEPII Working Paper Should everybody be in services? We retain four indicators of performance: protability (which we proxy by EBITDA  Earnings Before Interest, Taxes, Depreciation, and Amortization  divided by value added), 13 employment, turnover, and the production sales of goods. 4.1. Servitization premia Before we estimate the causal impact of servitization on performance, we provide evidence of the magnitude of the performance gap between servitized and non-servitized rms. To assess the dierences between the two groups of rms precisely, we have to remove possible composition eects and compare rms in the same year and industry. This is accomplished by estimating the following equation: P erf ormance = =49 X k i ;t k =0 ] ; +1] d] ; +1] k k k k ;i ;t 1 + Employment i ;t 1+ + ; where Performancei ;t is a variable characterizing the performance of rm j;t i (2) i ;t in year t: rm i 's protability (as a percentage), (log) employment, (log) turnover, and (log) sales of goods at time t .  is a 2-digit industryyear dummy, and  is the error term. The dummies d] ; +1] 1 are dened as follows: j;t k k ;i ;t d] ; +1] k k i ;t ;i ;t 1=  1 0 if k < Service Intensity  k + 1; k 2 [0%; 49%] d] ; +1] 1 takes the value 1 if the service intensity of ]k ; k + 1], where k varies from 0% to 49%.The coecients Each of the 50 dummy variables rm i ] ; +1] k k lies in the interval i ;t otherwise k k ;i ;t on these dummies are estimated, taking the performance of non-servitized rms as a reference. They are interpreted as the average performance gap (i.e., the premium) between pure manufacturers and rms with a given service intensity within the same year and industry. Because protability, turnover, and total sales of goods are likely to be correlated with rm size, we control for lagged employment in those regressions. 14 The results are presented in gures 6 and 7. We graphically report only the coecients ] ; +1] k k along with the 95% condence interval. Dark/plain circles represent signicant coecients while light/hollow circles represent coecients that are not statistically different from zero at the 95% level. It is noteworthy that around 87% of servitized rms in our sample are included in the rst ten dummies (i.e., services account for less than 10% of their production sales). In gure 6, we display the protability premia of being servitized rms. These premia are positive and statistically signicant, and they are remarkably stable over the range of service intensities. Regardless of the service intensity, servitized rms exhibit greater protability of 3.5 percentage points with respect to non-servitized rms of comparable size in their industry. In 2007, the median prot rate was 46%. An increase of 3.5 percentage points is equivalent to a 7.6% increase. The coecients 13 We ] ; +1] k k do not consider sales of products that are bought and sold without transformation by the rm. See Bernard and Fort (2013) on a description of factoryless goods producers, i.e. rms who do not produce themselves the goods they sell, but are involved in the design and coordinate their production. 14 The premia are very similar when we do not control for employment. 11 CEPII Working Paper Should everybody be in services? become non-signicant for rms with service intensities greater than 30%. Very few rms have a service intensity above 30% in our sample, which may explain the non-signicance of these coecients. Panel (a) of gure 7 shows the premia in terms of employment. The estimated coecients ] ; +1] are all positive and statistically signicant. k k Their magnitude decreases with service intensity, but they remain positive. In panel (b), we show how (the log of ) turnover of servitized rms compares with that of pure manufacturers. The results appear similar to those in panel (a). For service intensities below 30%, the premia are signicant, positive, and stable. For high levels of service intensity, the premia seem small, but the small number of rms in these categories sharply reduces the precision of the estimates. On average, servitized rms with a service intensity below 30% generate almost 20% more revenue than do non-servitized rms. In panel (c), we consider the sales of goods. The estimated coecients ] ; +1] are positive and signicant for low levels of service intensity. k k They become negative and statistically signicant once the service intensity is greater than 20%. The positive and signicant coecients reveal that rms selling few services have larger sales of goods than rms that do not sell services at all. On average, rms with a service intensity below 10% sell 16% more goods than pure goods producers. Recall that most of the servitized rms in our sample (87%) have service intensity below 10%. The negative coecients on ] ; +1] when the service intensity is greater than 20% therefore k k concern very few rms. These results indicate a dual relationship between the production of services and the production of goods, which can be complements or substitutes. On the one hand, the provision of services is complementary to the production of goods when services represent a very small proportion of total rm production. On the other hand, some rms tend to specialize in the production of services, increasing their provision of services in lieu of goods production. 12 CEPII Working Paper Should everybody be in services? -10 -5 0 5 10 Figure 6  Relative Protability of Servitized Firms 0 10 20 30 % of services in total production 40 50 Figure 7  Relative Performance of Servitized Firms (employment, turnover, and production of goods) (b) Turnover (log) -.5 -.4 -.2 0 0 .5 .2 1 .4 (a) Employment (log) 20 30 % of services in total production 40 0 50 10 20 30 % of services in total production 0 .5 (c) Sales of goods (log) -.5 10 -1 0 0 10 20 30 % of services in total production 13 40 50 40 50 CEPII Working Paper 4.2. Should everybody be in services? The causal impact of servitization: empirical strategy The premia reported in gure 7 deliver two key messages. First, servitized rms have better performance than non-servitized rms: They are larger (in terms of employment and production) and more protable. Second, with the exception of the production of goods, service intensity does not inuence the premia much. Selling services is associated with better performance, even when they represent a very small share of rm sales. The premia do not increase with the share of services of total output. Together these results suggest that the decision to start selling services is what really matters, while the decision to sell more or fewer services does not seem to correlate with rm performance. Building on this observation, our causal analysis will focus on the decision to start selling services rather than on changes in the service intensity. 15 The premia shown above are simple OLS estimates and suer from patent endogeneity problems. Our rst concern is that some confounding factors could be simultaneously correlated with both the decision to start selling services and rm performance. The decision to start selling services may be motivated or inuenced by changes in rm environments (e.g., changes in competition pressure, technological changes, evolution of public regulations, improvement of transport and telecommunication infrastructures). The decision may also depend on unobserved rm-level characteristics, such as manager ability and past experiences. Failing to control for these confounding factors can seriously bias estimates. The second concern is reverse causality induced by self-selection. Do rms decide to sell services because they have good performance or do they have better performance because they also sell services to their consumers? The bias may occur in both directions. On the one hand, servitization may be a selective process whereby only the highest-performing rms nd it protable to sell services. This mechanism will be observed, for instance, if rms have to invest in and allocate some managerial resources to start selling services. These investments may not be aordable to rms with low prots or strong nancial constraints. They may also be non-protable for rms with low competitiveness because the potential commercial gain they expect from selling services may not compensate for the xed investment cost. In this case, the OLS estimates would be biased upward. On the other hand, a negative relationship between ex ante rm performance and the decision to start selling services may also exist. When facing a negative shock, rms may try to restore their market shares and prots by shifting their production toward services in order to generate additional revenues and/or to add value to the good they sell. If the decision to start selling services is a defensive strategy for declining rms, we would expect the 16 OLS estimates to be biased downward. Our response to these endogeneity problems is twofold. First, we control for unobserved confounding factors that may simultaneously inuence the decision to sell services and rm performance. The traditional method to address unobserved variables is to use rm- 15 In unreported regressions, we also examined how changes in the share of services of total sales aect rm performance. The estimations produced volatile and non-robust results, which conrms the ambiguous impact of service intensity on rm protability shown by Suarez et al. (2013). 16 Breinlich et al. (2014) provide empirical evidence in favor of such a defensive strategy. They show that increasing competition pressure resulting from lower European manufacturing taris caused British rms to shift into the provision of services and out of the production of goods. 14 CEPII Working Paper Should everybody be in services? level xed eects in a dierence-in-dierences approach. This is not the most appropriate method in our case. Dierence-in-dierences estimators are based on the assumption that the most relevant unobserved confounders are time-invariant, which may not be true here. It is very likely that rms that decide to sell services have recently experienced some specic shock: a negative shock that reduced their protability, a positive shock that provided them the resources needed to invest in a new activity, or simply a change in their management team or ownership structure that may inuence their strategies and performance in some undetermined way. 17 In this case, the most appropriate econometric strategy is to estimate an LDV model in which all relevant omitted variables (including those that are time varying) are controlled for by the lags of the dependent variable. Compared to a xed eects model, an LDV model oers better control for self-selection 18 and the ensuing reverse causality bias. In addition, we introduce year industry xed eects to capture unobserved determinants that may inuence performance in a given year and 2-digit industry (e.g., changes in technology, regulations, infrastructures, competitive environment). Our preferred specication is: P erf ormance = 1(serv 1 ) + i ;t #X Lags i ;t k k =1 P erf ormance where Performancei ;t measures the performance of rm the value one if the rm i sells services at t i k + 1+ + ; in year i ;t t j;t (3) i ;t (i.e., protability, 1 ) is a dummy variable taking 1 and zero otherwise;  1 is a vector of employment, turnover, or production of goods); 1(serv i ;t i ;t i ;t control variables, which are all lagged by one period to avoid simultaneity issues; 2-digit industry xed eects; and  a set of year i ;t  j;t is is the error term. As the accuracy of the parameter estimates tends to increase with the number of lags of the dependent variable (Wilkins, 2015), our preferred specication includes three lags (i.e., 19 The coecient of interest, in equation (3)). #Lags = 3 , measures the average treatment eect (ATE), i.e., the observed impact of sales of services on Performancei ;t . The LDV model, by explicitly controlling for the trend of past rm-level performance, addresses the omitted variables issue and helps alleviate concerns about reverse causality bias. Nevertheless, reverse causality bias may persist if rms start selling services because they anticipate changes in their performance. For instance, rms innovating in products 17 The literature has emphasized the role of organizational changes in successful transitions to services. When moving toward services, rms often need to change their organizational structures and business models. These changes are costly, and rms may fail to implement them successfully, thus leading to the previously describedservice paradox. Bowen et al. (1989) argue that managers in manufacturing companies are often reluctant to adopt service-specic values, as these values contradict traditional manufacturing goals and practices, such as standardization and eciency. This point is also made by Gebauer and Fleisch (2007), who argue that managers are highly risk-averse when it comes to replacing their traditional productoriented values with service-oriented values, a point also raised in Mathieu (2001) and Eggert et al. (2011). See Vargo and Lusch (2008) for a description of the goods-dominant and service-dominant logics in manufacturing rms. 18 As a robustness check, we also estimate a xed eects regression. The results, which are consistent with those obtained by the LDV model, are shown in appendix table 9. 19 Our results are robust to the use of only one or two lags. 15 CEPII Working Paper Should everybody be in services? 20 To identify a causal link between servitiza- may also decide to sell services with them. tion and rm performance, we need a suitable instrumental variable, that is, a measure correlated with the decision to start selling services but uncorrelated with the dependent variable. This is not an easy task with the data at hand, as any information on the rm's balance sheet is very likely to be correlated with its performance. Hence, we propose an instrument based on the assumption that management practice spillovers exist across rms. We consider that rms observe and imitate their competitors and are more likely to start 21 For each rm selling services if comparable rms in the neighborhood already do so. and year i t in our database, we compute the number of servitized rms in its industry and the decile of size (measured as the average number of workers over the period) weighted by the geographic distance to i. The distance between rms is the geodesic distance 22 For rms located in the between the cities in which the two headquarters are located. same city, we use a measure of the internal distance of the city equal to where A is the area of the city in km2 (Mayer and Zignago, 2011). (2=3) A=, p Hence, our instrument 23 We lag the instrument by two varies by year, 2-digit industry, city, and rm size decile. 24 periods, as our endogenous variable is the decision to sell services at . t 1 A legitimate concern about the instrument is that it may be directly correlated with rm performance. Indeed, if selling services inuences the competitiveness of rms, then changing the number of service suppliers in the neighborhood of a rm is very likely to alter the competition pressure it faces and thereby its performance. In this case, the exclusion restriction is not veried, and no inference can be drawn from the empirical results. Our empirical strategy addresses this issue in two ways. First, our instrument is lagged by two periods with respect to the dependent variable, Performancei ;t . This should eliminate simultaneity bias that would lead to violation of the exclusion restriction. Second, our rst- and second-stage regressions control for the past performance of rms (in t 2). t 1 and Any eect of our instrumental variable on past performance is therefore captured by these lags. It is quite unlikely that our instrumental variable inuences the current performance of rms without inuencing its performance in t 1 or t 2, a factor for which we explicitly control. Finally, we need to dene an appropriate control group given that our objective is to assess the impact of starting to sell services on rm performance. We compare the performance of rms that shift toward the provision of services to the performance of rms that do not (or have not yet started). In other words, we do not consider rms that sell services throughout the period. 20 Eggert We also omit rms that stop selling services to avoid mixing et al. (2011) and Visnjic et al. (2014) show that servitization is more likely to generate better performance when it is coupled with product innovation. 21 A vast empirical literature has shown that scanning the external environment to obtain information about competitors' practices is a determinant of management innovation at the rm level. See, for instance, Audretsch and Feldman (1996); McEvily and Zaheer (1999); Mol and Birkinshaw (2009) and Fu (2012). 22 The 2 French territory is divided into more than 36,500 cities with an average surface area of only 14.9 km . This high level of administrative fragmentation makes our measure of distance between rms quite precise and oers substantial variation in the instrument across rms. 23 We also performed robustness analyses with an alternative instrument: the share of servitized rms in the same industry located in the same or surrounding départements (France is divided into 95 départements). The (unreported) results are very close to those reported in the paper. 24 Our instrument is therefore lagged by only one period with respect to the endogenous variable.  rst-stage regression also includes all other explanatory variables and year 16 industry xed eects. The CEPII Working Paper Should everybody be in services? the eects of shifting into services from those of shifting out of services. Given these restrictions and the large number of lags (#Lags=3), the econometric identication relies on a sample of 6,392 individual rms and a total of 34,243 observations. Note that our results are robust to less restrictive alternative samples (see table 10 in the appendix). In the following, the average treatment eect (ATE), maximum likelihood with endogenous treatment, in equation (3), is estimated by full 1(serv i ;t 1 ). The rst-stage equation is a probit regression, which predicts the probability of treatment (i.e., the probability that a rm starts selling services). In all specications, the instrument provides a good t in the rst stage. The instrument has a signicantly positive impact on the probability of producing services and passes the usual validity tests. 4.3. Baseline econometric results 4.3.1. Protability We begin our presentation of the econometric results by examining the impact of servitization on the protability of rms in detail. The benchmark results are displayed in table 2. The coecient on of starting to sell services in the previous year. 1(serv i ;t 1 ) is the ATE Columns (1) and (4) show simple OLS estimates of the relationship between the lagged servitization dummy and the prot rate, excluding and including control for lagged employment, respectively. These specications, which do not take into account omitted variable or reverse causality issues, yield simple 25 The estimates premia of servitization that are comparable to those shown in gure 7. conrm the existence of a signicant premium. The protability of rms that start selling services is 4.2 percentage points higher than that of pure manufacturers (4.4 when controlling for lagged employment). In the sample used for the regressions, the average prot rate is 45.7%, which implies a premium on the prot rate between 9.2% and 9.6%, depending on whether we control for the number of employees. In columns (2) and (5), we control for the lagged values of the prot rate in order to account for selection eects. Unsurprisingly, the ATE decreases, conrming that most of the premium comes from self-selection. The impact of servitization remains positive but is much smaller. Starting to sell services is associated with an increase in the prot rate of between 0.4 and 0.47 percentage points. In columns (3) and (6), we properly control for endogeneity by instrumenting the decision to start selling services. The ATE is between 1.7 and 2.4 percentage points, which corresponds to a causal increase in the prot rate between 3.7 and 5.3%. The fact that the instrumental variable estimation provides a larger ATE reveals a negative endogeneity bias, which suggests that the shift toward the provision of services is driven by a quite strong defensive motive. It seems that, everything else equal, rms that start selling services anticipate a relative decline in their protability. Tables 9 and 10 in the appendix present a series of robustness checks. use alternative estimators and sets of controls. 25 The In table 9, we In line (1), we control for potential results shown in gure 7 are slightly dierent because they are based on sample of observations that is not restricted by the use of lagged variable or a precise control group. 17 CEPII Working Paper Should everybody be in services? determinants of protability: the lagged rm market share (the rm's sales of goods divided by the total sales of goods in the same industry) and the interaction between this market share and the lagged industry-level Herndahl index, which captures the level 26 The ATE is 2.46, which is not statistically dierent from competition in the industry. the benchmark regression (2.427, in column (6) of table 2). In lines (2)(4), we report the ATE obtained using a xed eect estimator instead of an LDV model. In line (2), we control only for the lagged employment level and do not instrument the treatment variable. Line (3) shows the ATE using an instrumental variable, and line (4) further controls for the rm's market share and its interaction with the Herndahl index. The ATEs are always positive and signicant, which conrms that starting to sell services boosts rm protability. Again, we observe that the instrumental variable estimates are substantially greater than the OLS estimates. Note also that the xed eects estimates are systematically larger than those of the LDV models. With xed eects, the ATE is approximately 4.6 (cf. lines (3) and (4) of table 9) compared to 2.4 with the LDV model (cf. column (6) of table 2). This is consistent with the bracketing properties of these two estimators, as described by Angrist and Pischke (2008). If the correct model is an LDV model, then xed eects will result in the overestimation of a positive treatment eect. However, if the most important omitted variables are time invariant, the correct model is a xed eects model, and the LDV estimator will result in an underestimation of the treatment eect. While we argue that the correct specication in our case is the LDV model, it is useful to think of our estimates as lower bounds of the true causal eect. In table 10, we test the robustness of our benchmark results to an alternative sample of rms. In our benchmark regression (table 2), we exclude rms that always sold services or that stopped selling services over the study period. We focused on the subsample of rms that either never sold services and that started to sell services. In line (1) of table 10, we exclude all rm that produced services during the year t-2. In other words, we use a sharper denition of the treatment. We exclude from the sample all treated rms once they have been treated (i.e., once they have started to produce services), and we estimate the impact of shifting toward services in the year after the shift but not in the following years. In line (2), we add to our benchmark sample rms that stop the production of services, and in line (3), we replicate our benchmark regression using the sample of rms that are always observed in our database (i.e., we exclude rms that appeared or disappeared between 1997 and 2007). We conrm that servitization has a positive inuence on rm protability. The point estimates obtained from these alternative samples are slightly larger than that reported in table 2. This further conrms that our preferred result is a conservative estimate of the impact of servitization on rm protability. 4.3.2. Employment, turnover and production of goods We now turn to the inspection of the impact of servitization on alternative indicators of rm performance. We re-estimate equation (3) using level of employment, turnover and production of goods as alternative dependent variables. The baseline results are shown in table 3. As a robustness check, we report the xed eects estimates in table 11. The results conrm the positive impact of servitization. Once again, the xed eects estimates 26 We do not include the Herndahl index on its own because it is fully captured by the industry eect. 18  year xed CEPII Working Paper Should everybody be in services? Table 2  Impact of Servitization on Firm Protability  Benchmark Results Dep. var.: Proti ;t Estimator: 1(serv 1) i ;t (1) (2) (3) (4) (5) (6) OLS LDV LDV-IV OLS LDV LDV-IV a (0.284) Proti ;t a 4.179 0.468 (0.109) 1 a 0.601 (0.011) Proti ;t 2 a 0.183 (0.010) Proti ;t 3 a 0.127 (0.008) Ln Employmenti ;t 2 34,243 34,243 0.014 0.740 Wald test (p-value)  All regression include industry  theses are clustered by industry Coecients on 4.418 (0.322) 0.599 1(ser v i ;t a 1 ) year. 0.399 (0.374) 0.601 (0.011) 0.182 2.427 0.183 0.597 0.127 (0.008) 0.181 34,243 a (0.010) a 0.125 (0.008) c a (0.010) a (0.010) a a (0.578) a (0.011) a (0.010) 0.126 a (0.112) a -0.296 R Notes: a 1 # Obs.  1.669 a (0.008) b 0.084 -0.122 (0.151) (0.042) (0.078) 34,243 34,243 34,243 0.014 0.740 -0.0951 -0.156 (0.000) (0.000) year xed eects. Robust standard errors in paren- Signicance levels: c <0.1, : p b <0.05, : p are the average treatment eects (ATEs). a <0.01. : p Columns (3) and (6) report full maximum likelihood estimations, where the treatment variable, 1(ser v i ;t 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized rms in the corre-  sponding industry and size decile. The Wald test ( =0) is below 10%, indicating that we can reject the null hypothesis of no correlation between the treatment and the outcome errors. are a bit larger than those of the LDV (except for employment), which implies that the 27 ATE shown in table 3 is a conservative lower bound of the true causal eect. As for protability, we nd a signicant, positive causal impact of servitization on rm outcomes. Because each dependent variable is in logarithmic form, the magnitude of the estimated impact of starting the production of services is given by the exponent of the ATE. On average and relative to pure manufacturers, rms that start producing services exp(0:263)  = 1:30), turnover increase their level of employment by approximately 30% ( by 3.7%, and sales of goods by 3.6%. The magnitude of these causal eects might seem very large, especially the impact on the number of employees. However, one has to keep in mind that most rms in our sample are small businesses. The marginal eects estimated here apply to relatively small values. For instance, the median rm in our sample has no more than 9 employees. A 30% increase in the number of employees represents two additional jobs for the median rm. 4.4. Extensions All econometric results shown above point in the same direction. They conrm that servitization has a positive causal impact on the performance of manufacturing rms. These 27 For the sake of conciseness, we do not report all robustness analyses. We have also checked the robustness of these results to alternative models, samples of rms and control groups. All unreported estimates corroborate those shown in table 3. They are available from the authors upon request. 19 CEPII Working Paper Should everybody be in services? Table 3  Impact of Servitization on Firm Employment, Turnover, and Production of Goods  Lagged Dependent Variable Model Dep. variable (1) Employment std. err. # Obs.  (p-value) a (0.019) 34,243 -0.593 (0.000) a (0.009) 34,243 -0.069 (0.001) a (0.008) 34,243 -0.072 (0.000) ATE 0.263 (2) Turnover 0.036 (3) Prod. goods 0.035 Notes: industry  Lagged dependent variable models with 3 lags, controlling for year xed eects in all regressions and for lagged employment in (2) and (3). The treatment variable, 1(ser v i ;t 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry and size decile. and 3). industry Estimators: two-step (line 1) and full maximum likelihood (lines 2 The (std.  err.) column reports robust standard errors clustered by year. Signicance levels: c : p <0.1, b : p <0.05, a : p <0.01. econometric results contrast with the many case studies that highlight the diculties companies experience with reaping the benets of a servitization strategy (Gebauer et al., 2005; Martinez et al., 2010) However, our estimates are only average eects, which may hide heterogeneity by industry or rm type. In this section, we evaluate the consequence of starting to sell services on the long-run performance of various sub-samples of rms. These more detailed results indicate that the impact of servitization on rm performance is much less systematic than suggested by our benchmark regressions. 4.4.1. Long-run eects As emphasized in the literature, selling services is associated with long-term investments with consumers, and the benets of servitization may take time to materialize. Our baseline specication has considered rm performance in the year following a move toward services. Thus, our results may miss some of the long-run eects of servitization. In the following table, we present results for how rm performance is aected at t +3 28 by the move toward services. t + 1, t + 2 and The results indicate that the consequence of servitization on rm protability are spread over time. Our benchmark results indicate that in the year after a shift toward services, servitized rms increase their prot rate by 2.42 percentage points compared to pure manufacturers (see table 2). Table 4 shows that this gap grows steadily for at least three years, reaching 4.3 percentage points by year t + 3. The impact on employment is also persistent, but the dynamics are clearly dierent, as the eect decreases slowly over time. Four years after the switch, we still observe a signicant causal impact of servitization on exp(0:138) 1  = 14:8%), but this eect is half the size of that observed the year following the switch (exp(0:263) 1 = 30:1%). The impacts on the number of employees ( turnover and production of goods also fade over time but at faster rates. Four years after the shift toward services, the causal eect of servitization on production is very small and barely signicant, and the impact on the production of goods is no longer visible. All together, these results suggest that the supply of services does not really support the production of goods over the long run. Servitization seems to be mainly a strategy that leads rms to focus on their most protable activities and/or to dierentiate their 28 Because of data limitations, we cannot estimate the impact of servitization over a longer period. 20 CEPII Working Paper Should everybody be in services? products further in order to charge higher margins. Table 4  Impact of Servitization on Firm Performance  Long-run Eects std. err. # Obs.  (p-value) (1) 2.6246 (0.962) 27,851 -0.183 (0.015) (2) 3.204 a (1.239) 22,360 -0.232 (0.015) (3) 4.306 a (1.004) 17,486 -0.314 (0.000) (4) 0.194 a (0.020) 27,851 -0.458 (0.000) (5) 0.161 a (0.022) 22,360 -0.400 (0.000) (6) 0.138 a (0.024) 17,486 -0.351 (0.000) (7) 0.027 a (0.010) 27,851 -0.056 (0.013) (0.015) 22,360 -0.027 (0.466) ATE a (8) 0.013 c (9) 0.016 (0.009) 17,486 -0.037 (0.054) (10) 0.023 b (0.010) 27,851 -0.043 (0.052) (11) -0.002 (0.040) 22,360 0.013 (0.895) (12) 0.013 (0.009) 17,486 -0.025 (0.223) Notes: Dependent variable +1 +2 Prot +3 ln(Emp:) +1 ln(Emp:) +2 ln(Emp:) +3 Proti ;t Proti ;t i ;t i ;t i ;t i ;t ln(T urnover ) +1 ln(T urnover ) +2 ln(T urnover ) +3 ln(P rod:goods ) +1 ln(P rod:goods ) +2 ln(P rod:goods ) +3 i ;t i ;t i ;t i ;t i ;t i ;t Lagged dependent variable models with 3 lags, controlling for lagged employment (except for lines 46) and industry  year xed eects. The treatment variable, 1(ser v i ;t 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry and size decile. Estimators: full maximum likelihood (lines 13 and 712) and two-step (lines 46). Column (std. err.) reports robust standard errors clustered by industry   year. Column ( ) reports the value of  and the corresponding p-value of the Wald test of the exogeneity of the instrumented variables. Signicance levels: a : p 4.4.2. c <0.1, : p <0.01. b <0.05, : p Results by rm size In table 5, we examine whether the impact of servitization diers by rm size. We report the ATE estimated from samples of micro, small, and medium and large rms, as dened by the European Commission. 29 For each of the four performance measures, we observe quite strong eects for micro and small businesses. For micro rms, starting to produce services is associated with an increase in the prot rate of 2.8 percentage points (which corresponds to an average increase of approximately 6% given the average level of the prot rate for this group of rms). Starting to sell services also increases employment, turnover and production for these rms. The impact on employment is particularly large (25%) but represents a limited number of new jobs given the small size of these rms. In our sample, the 30 so the estimated impact of servitization median micro rm employs only 5 workers, corresponds to an increase of 1.25 workers in this category. For small rms, the impact on protability if larger. The ATE is more than 3.1 percentage points, which corresponds to an average increase in the prot rate of 7.2%. Employment grows by more than 38%, which represents slightly less than 7 new jobs in the median small rm. 29 Firms 31 are classied according to their average number of employees over the observation period. Micro rms have no more than 10 employees. Small rms have between 11 and 50 employees, and medium and large rms have more than 50 employees. 30 The 31 The average number of workers in this class of rm is very close to the median: 5.48. median number of workers in this category 18. 21 CEPII Working Paper Should everybody be in services? In contrast, we do not observe a signicant impact of servitization on the performance of medium and large rms (i.e., rms with more than 50 employees). This non-signicance may be due to the relatively small number of rms in this category (especially of switching rms), which reduces the precision of the estimates. This pattern may also indicate that rms that have managed to grow without feeling the need to produce services perform particularly well in the production of goods. Table 5  Impact of Servitization on Firm Protability, Employment, Turnover, and Production of Goods  By Firm Size Dep. variable ATE std. err. # Obs.  (p-value) Firm type (# employees) (1) Protability (2) Protability a (0.590) 18,643 -0.161 (0.000) micro (1-10) c (1.274) 12,519 -0.222 (0.043) small (11-50) 2.786 3.127 2.300 > 50) (3) Protability (1.632) 3,081 -0.196 (0.177) large ( (4) Employment 0.226 a (0.158) 18,643 -0.449 (0.000) micro (1-10) (5) Employment 0.327 a (0.030) 12,519 -0.823 (0.000) small(11-50) (6) Employment -0.038 (0.082) 3,081 -0.199(0.000) large ( a (0.011) 18,643 -0.068 (0.000) micro (1-10 a (0.021) 12,519 -0.185 (0.002) small(11-50) (7) Turnover 0.030 (8) Turnover 0.086 (9) Turnover (10) 0.022 Prod. goods (11) Prod. goods (12) Prod. goods Notes: 0.023 > 50) > 50) (0.126) 3,081 -0.082 (0.819) large( b (0.011) 18,643 -0.063 (0.000) micro (1-10 a (0.017) 12,519 -0.180 (0.000) small(11-50) (0.123) 3,081 -0.467 (0.185) 0.083 0.168 > 50) large( Lagged dependent variable models with 3 lags, controlling for lagged employment and industry  year xed eects. The treatment variable is instrumented by the (2-year lagged) distance- weighted sum of servitized producers in the corresponding industry and size decile. Estimators: Full maximum likelihood (lines 13 and 712) and two-step (lines 46). Column (2) reports robust standard errors clustered by industry-year. Column (4) reports the value of  P-value of the Wald test of the exogeneity of the instrumented variables. <0.1, p 4.4.3. b : p <0.05, a : p and the corresponding Signicance levels: c : <0.01. Results by industry Moving toward supplying services to consumers is very likely to depend on the characteristics of the product being sold or on the type of competition prevailing within an industry. Using the same dataset, in Crozet and Milet (2014), we show that servitization is more widespread in industries that produce heterogeneous goods. 32 Fang et al. (2008) provide evidence that adding a service to the core product of the rm leads to higher market value, especially in industries with overall low growth and high volatility of sales. In a widely cited article, Teece (1986) argued that services do not loom large in the early stages of an industry. 32 We This inuenced the vision that services are benecial when rms use data on the rm-level exports of goods from French customs data and use Rauch (1999)'s classication to distinguish between dierentiated and homogenous goods. We nd a positive log-linear relationship between the average share of services of the industry's output and the share of dierentiated products in this industry. This argument is in line with Anderson et al. (1997), who argue that rm performance depends on the degree of heterogeneity of their product. In their paper, they link measures of productivity to customer satisfaction (which takes into account the standardization versus the customization quality of the product). 22 CEPII Working Paper Should everybody be in services? enjoy a well-established base of customers and as dierentiation of the product becomes increasingly dicult. This view has been recently challenged by authors, such as Suarez et al. (2013) and Cusumano et al. (2015), who note that services can be oered before, during, or after the purchase of the good. 33 To show how the impact of servitization diers by sector, we have assigned each rm in our database to a broadly dened sector 34 The results are reported in ta- and estimated equation (3) for each group separately. ble 6. Table 12 in the appendix shows the comparable results obtained for the sample of micro and small businesses. Again, this table conveys a more complex message about the consequence of servitization. At rst sight, this table conrms the positive impact of servitization on rm performance, as a very large majority of the estimates are signicantly positive (of the 28 estimates reported in this table, 16 are signicantly positive, 8 are non-signicant, and only 4 are negative). However, the results dier substantially across sectors and reveal both the complexity and the diversity of the servitization strategies discussed in the literature. We can identify three patterns: 1. Servitization has a positive impact on all four indicators of performance. This is the case for agri-food (1) and other manufacturing not elsewhere classied (7). For these two industries (particularly the latter group), the estimated impact is quite large but close to that of the pooled results. 2. Servitization increases the sales of goods (and more generally rm size) but not the prot rate. This is clearly the case for mineral and metal products (4) as well as for textile (2) (wherein turnover and sales of goods, but not employment, are positively aected). Here, sales of services are positively correlated with sales of goods. However, the service oering is not a strategy that increases protability, either because producing complementary services is not enough to provide a signicant competitive advantage or because the cost of organizing services activities outweighs (at least over the short-run) the competitive gains. 3. Servitization increases protability, but a substitution eect between services and goods prevails. This is the case for wood products, paper and printing (3) and chemicals and plastics (4). 35 The substitution eect observed suggests that the decision to supply services is a part of a broader strategy in which the rms focus on their most protable activities or markets. 33 A 36 very often-cited example is IBM, which introduced the rst computers for businesses in the 1950s. These products were expensive and unknown to consumers. IBM engaged in leasing contracts, which were combined with maintenance, and pay-for-usage contracts. Services preceded and substituted for the sale of these products. 34 All 35 For regressions control for 2-digit industry xed eects. the latter sector, the results are less clear than suggested by table 6. The negative impact on turnover and sales of goods shown in table 6 is entirely driven by the largest rms. Table 12 shows that for micro and small rms in this sector, servitization increases production. 36 Note that we nd no signicant impact in sector 6 (machinery, electrical, optical and transport equipment), except a small and very imprecisely estimated coecient for the number of employees. surprising result is mostly due to the behavior of the largest rms. This perhaps When we consider only micro and small rms, we obtain a positive and signicant impact of servitization on protability and employment (see table 12). 23 CEPII Working Paper Should everybody be in services? Table 6  Impact of Servitization  By Sector ATE std. err. # Obs.  (p-value) 1 Food, Beverage, Tobacco Protability Ln Employment 5.070 a (0.511) 10479 -0.390 (0.000) 0.178 a (0.034) 10479 -0.376 (0.000) b (0.013) 10479 0.214 (0.095) c (0.013) 10479 -0.066 (0.054) -1.268 (2.339) 1258 0.065 (0.629) Ln Turnover 0.026 Ln Prod. goods 0.024 2 Textile, Leather Protability Ln Employment Ln Turnover Ln Prod. goods 0.070 0.225 (0.116) 1258 -0.116 (0.662) b (0.073) 1258 0.357 (0.003) a (0.079) 1258 -0.373 (0.006) 0.214 3 Wood, Paper, Printing Protability 6.632 a (2.040) 4776 -0.453 (0.05) Ln Employment -0.080 (0.112) 4776 0.265 (0.365) a (0.031) 4776 0.745 (0.000) a (0.031) 4776 0.748 (0.000) Ln Turnover Ln Prod. goods -0.353 -0.357 4 Chemicals, Plastics Protability 9.690 a (0.987) 1527 -0.630 (0.000) Ln Employment -0.185 Ln Turnover Ln Prod. goods (0.256) 1527 0.584 (0.381) -0.237 b (0.093) 1527 0.564 (0.025) -0.248 b (0.097) 1527 0.566 (0.025) 2.860 (3.040) 9941 -0.173 (0.443) c (0.050) 9941 -0.217 (0.099) (0.019) 9941 -0.076 (0.106) (0.029) 9941 -0.035 (0.622) 5 Mineral, Metal Products Protability Ln Employment 0.096 Ln Turnover 0.044 b Ln Prod. goods 0.023 b 6 Machinery, Electrical Equip. Protability 1.220 (0.917) 4309 -0.028 (0.634) c (0.105) 4309 -0.462 (0.096) Ln Turnover 0.021 (0.019) 4309 -0.035 (0.163) Ln Prod. goods 0.021 (0.016) 4309 -0.044 (0.023) Ln Employment 0.190 7 Manufacturing, n.e.c. Protability a (1.942) 1953 9.085 (0.000) b (0.179) 1953 -0.749 (0.050) a (0.058) 1953 -0.191 (0.134) a (0.057) 1953 -0.178 (0.133) 7.997 Ln Employment 0.387 Ln Turnover 0.102 Ln Prod. goods 0.093 Notes: Lagged dependent variable model with 3 lags, controlling for lagged  employment and industry year xed eects. All lines report estimates by full maximum likelihood (except for employment: two steps), where the treatment variable, 1(ser v i ;t 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry and size decile. Column (std. err.) reports robust standard errors clustered by industry  year. Signicance levels: c <0.1, : p b <0.05, : p a <0.01. : p CEPII Working Paper 5. Should everybody be in services? Conclusion Servitization is growing everywhere, yet empirical evidence of its impact on rm performance remains scarce (Baines and Lightfoot, 2013). We contribute to lling this gap by documenting the extent and evolution of servitization in the French manufacturing sector between 1997 and 2007 using a large dataset of more than 50,000 servitized and non-servitized rms. We rst documented that the vast majority of French manufacturing rms report positive sales of services. While the share of servitized rms remained quite stable over the 19972007 period, we nd that the share of services of total production sales increased in all industries and, on average, in each rm. We showed that servitized rms are more protable, employ more workers, and have higher total sales than nonservitized rms. These premia depend greatly on whether rms sell services, but they do not vary with the share of services in production sales. Building on this result, we adopted a micro-econometric approach to assess the causal impact of engaging in the production of services on rm performance. We nd that, compared to rms that produce goods only, rms that start selling services increase their protability by 3.7% to 5.3%, increase their number of employees by 30%, and boost their sales of goods by 3.6%. From an academic perspective, several interesting questions that are beyond the scope of this paper are raised. Firms that complement their products with services have shifted toward a new business model. Their activities have become a mix of goods and services and no longer produce only tangible products. This raises the question of the relevance of unique industry classications based on the main activity of a rm. How should rms that produce as many goods as services be classied? On a more theoretical note, this paper raises the question of how to dene the production functions of such rms. It also calls into question the denition and proper calculation of the total factor productivity of these rms. References Acemoglu, D. and Guerrieri, V. (2008). Capital Deepening and Nonbalanced Economic Growth. Journal of Political Economy, 116(3):467498. Anderson, E. W., Fornell, C., and Rust, R. T. (1997). Customer Satisfaction, Productivity, and Protability: Dierences Between Goods and Services. Marketing Science, 16(2):129145. Angrist, J. and Pischke, J.-S. (2008). Mostly Harmless Econometrics: An Empiricist's Companion. Princeton University Press. Audretsch, D. B. and Feldman, M. P. (1996). R&D Spillovers and the Geography of Innovation and Production. American Economic Review, 86(3):63040. Baines, T. and Lightfoot, H. (2013). Made to Serve: How Manufacturers Can Compete Through Servitization and Product Service Systems. Wiley. Baines, T., Lightfoot, H., Benedettini, O., and Kay, J. (2009). The Servitization of Manufacturing: A review of literature and reection on future challenges. Journal of Manufacturing Technology Management, 20(5):547567. Baumol, W. J. (1967). Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis. American Economic Review, 57(3):415426. 25 CEPII Working Paper Should everybody be in services? Benedettini, O., Clegg, B., Kafouros, M., and Neely, A. (2010). The Ten Myths of Manufacturing. What Does the Future Hold for UK Manufacturing? Report prepared for the engineering and physical science research council. Benedettini, O., Swink, M., and Neely, A. (2013). Firm's Characteristics and Servitization Performance: A Bankruptcy Perspective. University of cambridge, cambridge service alliance working paper. Bernard, A. B. and Fort, T. C. (2013). Factoryless Goods Producers in the U.S. NBER Working Papers 19396. Bowen, D. E., Siehl, C., and Schneider, B. (1989). A Framework for Analyzing Cus- tomer Service Orientations in Manufacturing. The Academy of Management Review, 14(1):pp. 7595. Breinlich, H., Soderbery, A., and Wright, G. C. (2014). From Selling Goods to Selling Services: Firm Responses to Trade Liberalization. C.E.P.R. Discussion Papers 10116. Crozet, M. and Milet, E. (2014). The Servitization of French Manufacturing Firms. CEPII Working Paper 2014-10. Cusumano, M. A., Kahl, S. J., and Suarez, F. F. (2015). and the Competitive Strategies of Product Firms. Services, Industry Evolution, Strategic Management Journal, 36(4):559575. Eggert, A., Hogreve, J., Ulaga, W., and Muenkho, E. (2011). Product Innovations, and Firm Protability: Industrial Services, A Multiple-group Latent Growth Curve Analysis. Industrial Marketing Management, 40(5):661  670. European Commission, T. (2014). For a European Industrial Renaissance, Communication from the Commission to the European Parliament, The Council, The European Economic and Social Committee and the Committee of the Regions. Fang, E., Palmatier, W. R., and Steenkamp, J.-B. E. (2008). Eect of Service Transition Strategy on Firm Value. Journal of Marketing, 72(5):114. Fu, X. (2012). Foreign Direct Investment and Managerial Knowledge Spillovers through the Diusion of Management Practices. Journal of Management Studies, 49(5):970 999. Fuchs, V. R. (1968). The Service Economy. Number 87. NBER. Gebauer, H. and Fleisch, E. (2007). An Investigation of the Relationship Between Behavioral Processes, Motivation, Investments in the Service Business and Service Revenue. Industrial Marketing Management, 36(3):337  348. Gebauer, H., Fleisch, E., and Friedli, T. (2005). Overcoming the Service Paradox in Manufacturing Companies. European Management Journal, 23(1):1426. Lodefalk, M. (2013). Servicication of Manufacturing - Evidence from Sweden. International Journal of Economics and Business Research, 6(1):87113. Martinez, V., Bastl, M., Kingston, J., and Evans, S. (2010). Challenges in Transforming Manufacturing Organisations into Product-service Providers. Journal of Manufacturing Technology Management, 21(4):449469. Mathieu, V. (2001). Service strategies within the manufacturing sector: benets, costs and partnership. International Journal of Service Industry Management, 12(5):451475. Mayer, T. and Zignago, S. (2011). Notes on CEPII's Distances Measures: The GeoDist 26 CEPII Working Paper Should everybody be in services? Database. CEPII Working Papers 2011-25. McEvily, B. and Zaheer, A. (1999). Bridging Ties: A Source of Firm Heterogeneity in Competitive Capabilities. Strategic Management Journal, 20(12):11331156. Mol, M. J. and Birkinshaw, J. (2009). The Sources of Management Innovation: When Firms Introduce New Management Practices. Journal of Business Research, 62(12):12691280. Neely, A. (2008). Exploring the Financial Consequences of the Servitization of Manufacturing. Operations Management Research, 1(2):103118. Neely, A., Benedittini, O., and Visnjic, I. (2011). The Servitization of Manufacturing: Further Evidence. Euoma conference, cambridge. Ngai, L. R. and Pissarides, C. A. (2007). Structural Change in a Multisector Model of Growth. American Economic Review, 97(1):429443. Oliva, R. and Kallenberg, R. (2003). Managing the Transition from Products to Services. International Journal of Service Industry Management, 14(2):160172. Pilat, D., Cimper, A., Olsen, K., and Webb, C. (2006). The Changing Nature of Manufacturing in OECD Economies. STI Working Paper 9. Rauch, J. E. (1999). Networks Versus Markets in International Trade. Journal of International Economics, 48(1):735. Reinartz, W. and Ulaga, W. (2008). How to Sell Services More Protably. Harvard Business Review, 86(5):9098. Stigler (1956). Trends in Employment in the Service Industries. National Bureau of Economic Research, General Series n.59. Suarez, F. F., Cusumano, M. A., and Kahl, S. J. (2013). Services and the Business Models of Product Firms: An Empirical Analysis of the Software Industry. Management Science, 59(2):420435. Teece, D. J. (1986). Proting from Technological Innovation: Implications for Integration, Collaboration, Licensing and Public Policy. Research Policy, 15(6):285305. Vandermerwe, S. and Rada, J. (1988). Servitization of Business: Adding Value by Adding Services. European Management Journal, 6(4):314324. Vargo, S. L. and Lusch, R. F. (2008). From Goods to Service(s): Divergences and Convergences of Logics. Industrial Marketing Management, 37(3):254  259. Visnjic, I., Wiengarten, F., and Neely, A. (2014). Only the Brave: Product Innovation, Service Business Model Innovation, and Their Impact on Performance. Journal of Product Innovation Management, page DOI: 10.1111/jpim.12254. Wilkins, A. S. (2015). To Lag or Not to Lag? Re-evaluating the Use of Lagged Dependent Variables in Regression Analysis. mimeo. Wise, R. and Baumgartner, P. (1999). Go Downstream in Manufacturing. Business Review, 135. 27 Harvard CEPII Working Paper 6. Should everybody be in services? Appendix 6.1. Summary statistics  complete sample Table 7  Summary Statistics - Sample including Firms with a Share of Services of Total Sales above 50% All rms (1) # Firms Servitized rms 1997 2007
97-07 1997 2007
97-07 31,603 28,258 -1.11 23,345 21,007 -1.02 73.9 74.6 1,677 1,473 88.0 91.0 Share (%) (2) Employment Total (thousand) 1,905 1,618 -1.62 Share (%) Average (3) Turnover e, Total ( million) 60.27 57.27 -0.51 71.8 69.7 -0.28 333.1 415.8 +2.24 298.8 387.6 +2.64 89.7 93.2 Share (%) Average ( (4) e, thousand) Production of goods e, Total ( million) 10.5 14.7 +3.39 12.8 18.4 +3.69 281.0 334.6 +1.76 247.6 307.0 +2.17 88.1 91.7 Share (%) Average ( (5) (6) e, 8.9 11.8 +2.91 10.6 14.6 +3.22 Average (%) 47.0 48.7 +0.35 47.7 49.8 +0.43 Average (%) 17.0 18.3 +0.72 23.1 24.5 +0.62 Median (%) 1.0 1.3 +2.63 2.6 3.2 +2.05 0.37 0.37 Protability thousand) -1.29 Service intensity Std. dev.
97-07 corresponds to the annualized growth rate between 1997 and 2007. This sample of rms produces mainly goods (i.e., whose service intensity is below 50% over the period) or mainly services. Servitized rms are rms reporting strictly positive sales of services. 28 CEPII Working Paper 6.2. Should everybody be in services? Transition matrix Table 8  Transition Matrix - Change in Service Intensity between 1997 and 2007  from to 0% bin1 bin2 bin3 bin4 bin5 bin6 bin7 bin8 bin9 bin10 0% 11.98 7.42 0.40 0.17 0.11 0.06 0.05 0.05 0.05 0.04 0.25 5.89 39.49 2.78 0.82 0.40 0.17 0.15 0.10 0.06 0.08 0.42 bin1 bin2 0.23 1.79 1.17 0.47 0.18 0.12 0.05 0.04 0.01 0.04 0.03 bin3 0.09 0.63 0.45 0.42 0.20 0.14 0.07 0.03 0.02 0.03 0.05 0.04 0.26 0.13 0.21 0.19 0.12 0.11 0.07 0.02 0.01 0.05 bin4 bin5 0.03 0.13 0.07 0.11 0.13 0.18 0.08 0.07 0.07 0.01 0.04 bin6 0.01 0.11 0.03 0.06 0.05 0.11 0.12 0.08 0.04 0.04 0.07 bin7 0.01 0.06 0.03 0.04 0.04 0.05 0.08 0.10 0.08 0.08 0.05 bin8 0.02 0.08 0.01 0.02 0.03 0.03 0.04 0.05 0.12 0.09 0.12 bin9 0.02 0.07 0.01 0.01 0.01 0.02 0.03 0.04 0.08 0.14 0.23 0.08 0.31 0.06 0.02 0.02 0.03 0.03 0.05 0.08 0.13 2.36 bin10 100% 0.83 1.38 0.22 0.18 0.12 0.12 0.14 0.13 0.13 0.15 1.63 Notes: Constant sample of 29,909 rms. Lines refer to the service intensity in 1997, while columns refer to the service intensity in 2007. intensity. 100% 1.30 1.57 0.24 0.13 0.13 0.09 0.08 0.10 0.16 0.20 1.96 3.81 Bins are dened as 10% intervals of service Firms in bin5 have a service intensity between 40% and 50%. The rst and last columns (0% and 100%) refer to rms that produced either only goods or only services, respectively, in 2007. 6.3. Robustness checks Table 9  Impact of Servitization on Firm Protability  Alternative Controls and Estimators std. err. # Obs.  (p-value) Method a (0.592) 34,243 -0.158 (0.001) LDV-IV a (0.165) 34,243 - FE a (0.669) 34,243 -0.152 (0.000) FE -IV a (0.665) 34,243 -0.153 (0.000) FE-IV ATE (1) (2) 2.459 1.693 (3) 4.585 (4) 4.597 Notes: Stoppers and continuously servitized rms are excluded. reports the robust standard errors clustered by industry value of  Controls 3 1 1 2 Column (std. err.) year. Column (4) reports the  and the corresponding P-value of the Wald test. P-values below 10% indicate that we can reject the null hypothesis of no correlation between the treatment errors 1 = lagged 2 = 1 , lagged market share and interaction between lagged market share and lagged industry-level Herndahl index; 3 = 2 , Prot 1 , Prot 2 and and the outcome errors. Column (6) indicates the set of control variables: employment level; i ;t Proti ;t 3. i ;t Line (1) reports lagged dependent variable model estimates. Lines (2), (3) and (4) report rm-level xed eects estimates. Lines (1), (3) and (4) report the full maximum likelihood estimates of the ATE, where the treatment variable, 1(ser v i ;t 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry. Line (2) reports the OLS estimates. All regressions include industry  year xed eects. Signicance levels: c 29 <0.1, : p b <0.05, : p a <0.01. : p CEPII Working Paper Should everybody be in services? Table 10  Impact of Servitization on Firm Protability  Alternative Samples ATE std. err. (0.878) 27,415 -0.123 (0.018) c (1.834) 72,034 -0.207 (0.154) a (0.899) 22,304 -0.213 (0.003) b (1.375) 17,300 -0.171 (0.045) a (1.499) 51,232 -0.421 (0.000) (1) 2.575 2.953 (4) (5) Samples a (2) (3)  (p-value) # Obs. 3.168 3.457 5.803 Starts With All only stops years X X X X X X X Notes: Lagged dependent variable model with 3 lags, controlling for lagged employment and industry  year xed eects. treatment variable, 1(ser v i ;t All lines report estimates by full maximum likelihood, where the 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry. Column (std. err.) reports robust stan-  dard errors clustered by industry year. Column (4) reports the value of  and the corresponding P-value of the Wald test of the exogeneity of the instrumented variables. Line (1) uses the 2. t sample of rms that were not producing services at those that always produce services. Line (2) includes all rms except Line (3) replicates the benchmark regressions shown in column (6) of table 2 for the panel of rms active from 1997 to 2007. Lines (4) and (5) replicate the regressions shown in lines (1) and (2) for the panel of rms active from 1997 to 2007. Signicance levels: c <0.1, : p b <0.05, : p a <0.01. : p Table 11  Impact of Servitization on Firm Employment, Turnover, and Production of Goods  Fixed Eects Estimates Variable (1) (2) (3) Employment Turnover Prod. goods std. err. # Obs.  (p-value) a (0.032) 34,243 0.026 (0.7418) a (0.014) 34,243 -0.093 (0.000) a (0.014) 34,243 -0.095 (0.000) ATE 0.068 0.116 0.087 Notes: Fixed eects model controlling for industry  year xed eects in all re- gressions and for lagged employment in (2), (3), and (4). The treatment variable is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry and size decile. Estimator: two-step (line 1) and full maximum likelihood (lines 2 and 3). The (std. err.) column reports robust standard errors clustered by industry value of   year. The last column reports the and the corresponding P-value of the Wald test of the exogeneity of the instrumented variables. Signicance levels: 30 c <0.1, : p b <0.05, : p a <0.01. : p CEPII Working Paper Should everybody be in services? Table 12  Impact of Servitization by Sector  Micro and Small Businesses ATE std. err. # Obs.  (p-value) 1 Food, Beverage, Tobacco a (0.437) 9952 -0.375 (0.000) a (0.028) 9952 -0.337 (0.000) a (0.012) 9952 -.101 (0.001) a (0.013) 9952 -0.102 (0.000) Protability -0.920 (2.427) 1105 0.141 (0.848) Ln Employment 0.118 Protability Ln Employment 4.879 0.156 Ln Turnover 0.040 Ln Prod. goods 0.039 2 Textile, Leather (0.135) 1105 -0.195 (0.103) a (0.064) 1105 -0.382 (0.001) a (0.073) 1105 -0.362 (0.002) (1.970) 4430 -0.458 (0.004) Ln Turnover 0.234 Ln Prod. goods 0.232 3 Wood, Paper, Printing Protability 6.889 a a Ln Employment 0.265 (0.054) 4430 -0.576 (0.000) Ln Turnover -0.363 (0.033) 4430 0.750 (0.000) 0.037 (0.047) 4430 -0.084 (0.428) 10.012 (1.392) 1187 -0.649 (0.000) 0.128 (0.129) 1187 -0.240 (0.501) b (0.121) 1187 -0.559 (0.070) c (0.131) 1187 -0.517 (0.114) Ln Prod. goods a 4 Chemicals, Plastics Protability Ln Employment Ln Turnover Ln Prod. goods a 0.252 0.233 5 Mineral, Metal Products Protability Ln Employment Ln Turnover Ln Prod. goods 3.334 (2.598) 9216 -0.200 (0.298) a (0.030) 9216 -0.575 (0.000) a (0.009) 9216 -0.120 (0.000) a (0.010) 9216 -0.108 (0.000) 0.236 0.065 0.056 6 Machinery, Electrical Equip. Protability Ln Employment 2.759 c (1.545) 3601 -0.120 (0.249) a (0.042) 3601 -0.358 (0.000) 0.151 Ln Turnover 0.030 (0.023) 3601 -0.054 (0.122) Ln Prod. goods 0.028 (0.022) 3601 -0.059 (0.060) (2.074) 1671 -0.566 (0.000) 7 Manufacturing, n.e.c. Protability Ln Employment Ln Turnover Ln Prod. goods 8.348 a a 0.417 0.118 b 0.124 (0.155) 1671 -0.771 (0.015) (0.072) 1671 -0.234 (0.107) (0.061) 1671 -0.257 (0.031) Notes: Lagged dependent variable model with 3 lags, controlling for lagged  employment and industry year xed eects. All lines report estimates by full maximum likelihood (except for employment: two-step estimator), where the treatment variable, 1(ser v i ;t 1 ), is instrumented by the (2-year lagged) distance-weighted sum of servitized producers in the corresponding industry and decile of size. Column (std. errors clustered by industry year. The last column reports the value of  err.) reports robust standard  and the corresponding P-value of the Wald test of the exogeneity of the instrumented variables. Signicance levels: c : p 31 <0.1, b : p <0.05, a : p <0.01.