More and Better Jobs,But Not for Everyone: Effects of Innovation in French Firms

Innovation and Employment

In standard economic theory, technological innovation has ambiguous employment effects at the firm level that will also depend on the type of innovation (product or process) (Van Reenen 1997). Process innovation may decrease the level of employment through a direct labor-saving effect: The required level of employment for a given output decreases when a firm implements a new production process. Compensation mechanisms might mitigate or even overcome that negative impact, however, because such process innovations also reduce the effective cost of labor and may lead firms to increase output. Product innovation leads to the opening of new markets or to an increase in the range or quality of products or services, which should have a job creation effect. This effect should be stronger when the innovation is more radical, for instance, when the firm is the first to implement an innovation at the international or market level. Yet, even in the case of new products (goods or services), some contradictory mechanisms are at play; for example, new products might displace the older ones, reducing the positive effect on employment. Much empirical literature (see Vivarelli 2014 and Calvino and Virgillito 2018 for very detailed reviews) concludes somewhat positive employment effects of product innovation in various countries in recent years, especially when combined with patent applications and when the product innovation takes place in high-technology firms. Effects are mixed for process innovation. The literature shows often insignificant effects and sometimes negative effects in line with the labor-saving hypothesis at the firm level (Calvino and Virgillito 2018). We herein follow this literature but also more specifically focus on two issues on which research has, to date, been more limited: the effects of innovation on some dimensions of job quality and the potential heterogeneity in the employment and job-quality effects, by both occupation and industry.

Innovation and Job Quality

Considering a multidimensional approach to job quality as in recent socioeconomic literature (Davoine, Erhel, and Guergoat-Larivière 2008; Muñoz de Bustillo, Fernández-Macías, Esteve, and Antón 2011; Green et al. 2013; Osterman 2013), we can make a number of hypotheses about the general impact of innovation on job quality. First, technological change is a determinant of productivity, which is generally thought to be positively correlated with job quality. This effect, however, depends on workers’ bargaining power and their ability to capture the returns from higher productivity. Second, technological change can affect the structure of the workforce, leading to the creation and/or destruction of good and bad jobs and distinct aggregate job qualities (at the macro but also the firm level). Finally, the adoption of new technologies may have significant effects on the work environment, work organization, task division, and working conditions (Muñoz de Bustillo, Grande, and Fernández-Macías 2016). We herein focus more specifically on employment-quality variables—namely, wages, employment contracts, and working hours—and identify the precise channels through which innovation affects workers. These variables do not reflect the whole set of dimensions that may be included in a job quality concept (see, e.g., Gallie 2007; Green et al. 2013), but their importance for workers is confirmed by empirical analyses focusing on the determinants of job satisfaction. Indeed, the literature on job satisfaction has shown that earnings, job security, and working hours matter for job satisfaction (see, e.g., Clark 2005 or Sousa-Poza and Sousa-Poza 2000 on International Social Survey Programme data). In the French case, higher wages and higher working hours increase job satisfaction, and temporary contracts tend to decrease job satisfaction (Davoine 2007).

From a theoretical point of view, the effects of innovation on wages are ambiguous (Calvino and Virgillito 2018). In the neoclassic approach, workforce displacement after innovation leads to a greater labor supply followed by a decrease in wages that produces the labor demand required to put the labor market back in equilibrium. In a Keynesian-Schumpeterian perspective, however, higher productivity that follows innovation can produce an increase in wages if workers are able to appropriate some of the productivity gains. More generally, the bargaining power of workers may affect the relationship between innovation and the various dimensions of job quality.

Considering the type of employment contract, we expect the effect of innovation—which is generally found to be positive on firm-level employment—to favor permanent employment if the firm expects the innovation to have durable positive consequences on its activity and has a long-run strategy (labor-force hoarding and investment in human capital and specific skills). By contrast, innovation can encourage temporary contracts if it is, instead, considered to be more transitory and occurring in depressed economic conditions (Malgarini, Mancini, and Pacelli 2013), again, depending on the firm’s strategy (Antonucci and Pianta 2002).

Last, the effect of technological change and especially digitization on the number of hours worked is also ambiguous. Productivity may increase and lead to a decrease in working hours, as has been seen in developed countries over the past century. Conversely, from a frictional labor-market perspective (with a scarcity of specific skills), innovation may increase working hours for some categories of workers. Digitization can also bring greater potential to work longer hours away from the workplace or from home. In terms of job quality, longer working hours can be considered an improvement for some workers (e.g., involuntary part-timers) but can also mean excessive working hours for full-timers.

Only a few empirical economic analyses have considered the effects of innovation on job quality. Those that have mostly focus on wages and investigate the impact of innovation on the relative dynamics of wages and not on the absolute levels of wages. At the firm level, only a few studies exist (mainly on the United Kingdom or the United States), and they show contrasting effects of innovation on wages, either insignificant or positive (Van Reenen 1996; Doms, Dunne, and Troske 1997; Aghion, Bergeaud, Blundell, and Griffith 2017). Even less literature has also looked at the relationship between the type of employment contract and innovation, mostly from the point of view of the effect of the former on the latter, and not the potential effect of innovation on the development of permanent and/or temporary jobs. Using data from various countries, most of these past studies reported a negative relationship between the use of flexible contracts and innovation. Firms with greater shares of temporary workers have fewer patents (Franceschi and Mariani 2016), lower sales of innovative new products (Zhou, Dekker, and Kleinknecht 2011), and lower R&D investment (Vergeer and Kleinknecht 2014). Only Giuliodori and Stucchi (2012) concluded that product and process innovation favored both temporary and permanent employment following the decrease in employment protection legislation in Spain in 1997 (previously, innovation had increased only temporary employment).

Heterogeneous Effects of Innovation by Occupation and Industry

We also question the potentially heterogeneous effects of innovation by occupation. In the economic literature, under skill-biased technological change, innovation favors higher-skilled employment and destroys low-skilled jobs. Considerable empirical support now exists for this hypothesis using national, sectoral, and firm-level data (Autor, Katz, and Krueger 1998; Machin and Van Reenen 1998). More recently, however, the job-polarization hypothesis has appeared in both the economic literature and political debate, with a number of empirical contributions (Goos et al. 2014; Autor 2015; Eurofound 2015). This hypothesis describes the process by which low- and high-skilled jobs are simultaneously created in most economies, while middle-skilled jobs disappear. Although the level of analysis in the empirical polarization literature (the aggregate employment level) differs from that mentioned previously (firm-level data), it does seem crucial to evaluate the effect of technological innovation separately for specific groups of workers, especially by skill level. We thus consider potentially diverse effects by occupation, which has rarely been done at the firm level (Calvino and Virgillito 2018).

We also carry out estimations by industry to see if the effects differ in manufacturing and services. The existing results for the employment effects of innovation generally come from the manufacturing sector. Very few studies have focused on the effect of technological innovation in services, even though services account for more than 75% of private employment in France (and in many developed countries). The scarcity of analyses on innovation in services is because technological innovation is historically associated with manufacturing. Various types of approaches coexist about how technological innovation should be understood and analyzed in services: Assimilation (or technologist) approaches argue that theories and concepts developed for manufacturing apply in services; demarcation approaches stand that specific theories and tools are required to analyze innovation in services; and synthesis approaches try to think about a comprehensive framework that would borrow from each approach to analyze innovation across the economy (Coombs and Miles 2000; Gallouj and Savona 2010). The few existing quantitative analyses of innovation in services usually adopt the first perspective, using the same tools and hypotheses for services as for manufacturing. Certain characteristics of services lead us to question the transposability of hypotheses on employment effects of innovation from manufacturing to services, however. First, the degree of competition is generally expected to be lower in services than in manufacturing. In manufacturing, product innovation is a way to challenge competitors, yet it may be less the case in services because of lower competition. Second, some authors also point out the intangibility of the service product, which makes it difficult to convince consumers about the superiority of innovative services (Miles 2010) and could lead to lower effects of product innovation on sales and employment in services. Combined with the hypothesis on the lower level of competition, even when sales increase, this may not lead to higher employment if it only increases the profit margin. Third, the diversity of firms in services is even stronger than in manufacturing. Services include very diverse activities—from microbusinesses in family shops, consultants, accountants, and so on, to very large organizations in finance or insurance—so the overall effect of innovation in services may be a sum of diverse and heterogeneous effects. In empirical quantitative approaches, services are usually analyzed as a whole and the effects of technological innovation on employment are found to be lower than in manufacturing (Ugur, Churchill, and Solomon 2018).

A Firm-Level Database Linking Innovation and Employment Outcomes

We use three separate databases at the firm level: the Community Innovation Survey (CIS), administrative data on employment, and fiscal data.

The CIS was designed at the European level to collect data on innovation activities in firms following the Oslo Manual definitions of innovation (see text Appendix). In France, the sample is 18,109 firms in the market sector in CIS 2014 (after nonresponse and unusable questionnaires are dropped). ¹ The French survey data are considered to be of good quality, with an unweighted nonresponse rate of 25%, which is far below the nonresponse rates observed in many other European countries (for instance, 49% in Germany and 35% in Italy). ² The database includes only firms with 10 or more employees and is exhaustive for firms with more than 250 employees. It covers most market activities classified by the Nomenclature of Economic Activities (NACE, sections B to N), including the following industries: mining, manufacturing, electricity, water supply, construction, retail, transport, hotels and restaurants, information and communication, financial and insurance services, real estate, technical, scientific activities and administrative services. The industries excluded are agriculture, public services, education, health, arts and entertainment, extraterritorial activities, and private household employment (such as babysitting and cleaning).

The DADS (Déclarations Annuelles de Données Sociales) are administrative data on employment collected every year on the basis of establishments’ compulsory declarations. These include information collected at the establishment level in the private sector on employment, as well as contract type (fixed-term or permanent), annual working hours, and wages. Although employment, working hours, and wages can be disaggregated by occupation, contract types (fixed-term or permanent) cannot. These administrative data are of very good quality, with their main limitation being only the small number of variables that are collected (in particular, they contain no information on the work environment). We aggregate this establishment-level data up to the firm level to match them with the other data sets.

Our fiscal data (FARE-FICUS) include the standard accounting data used by the government to collect taxes on benefits and other assets. These databases provide information on productivity and labor costs. ³

We construct our database by merging CIS, DADS, and FARE-FICUS at the firm level, which yields a sample of 14,491 firms. The coverage of the merged data set is the same as CIS coverage in terms of firm size and industry (firms with 10 or more employees and exhaustive for firms with more than 250 employees, covering the NACE B to N sectors). The final sample represents 17% of total employment in France in 2011 and 28% of the total value added. ⁴

The merged database includes three sets of variables. The first concerns technological innovation behavior (as declared by firms in the 2012 to 2014 CIS data), separated into product and process innovation in accordance with the Oslo Manual typology and the CIS questionnaire (see text Appendix). We capture the intensity of product innovation by two complementary measures of innovation: product innovation that is “new to the market” and product innovation that is accompanied by a patent application. Although organizational innovation is also identified in the CIS, we do not consider it here as it may reflect heterogeneous management choices for which the potential job-quality effects are not straightforward (Rubery and Grimshaw 2001; Lam 2004).

The second set of variables covers the employment and job-quality outcomes in firms (available every year). This includes the total number of firm employees, the number of employees decomposed by occupation, and three dimensions of job quality: hourly wages, average annual number of hours worked per employee, and type of employment contract (permanent or fixed-term). In comparison to the job-quality literature, our approach focuses on employment quality and does not include the work environment or working conditions, for which there is no comprehensive database that could be matched to our data for France. We interpret higher wages and more permanent contracts as reflecting greater job quality (in accordance with job satisfaction analyses, as previously mentioned). The interpretation for working hours is less straightforward, as we cannot disentangle voluntary from involuntary increases in hours worked. In the French context, however, considerable involuntary part-time employment is present and overtime is quite heavily regulated, suggesting a positive relationship between hours of work and job quality. The third set of variables (available every year) relates to firms’ structural characteristics and economic performances (industry, size, age, productivity, labor costs, and so forth). We use two sectoral classifications: the usual one distinguishing among manufacturing, construction, retail, and services (based on NACE codes), and another developed by Eurostat to better characterize firms by their levels of technology and knowledge intensity (distinguishing among high-tech, medium high-tech, medium low-tech, and low-tech manufacturing, and between knowledge-intensive and less knowledge-intensive services). ⁵

This merged database then tells us whether the firm innovated between 2012 and 2014 and provides information on firm employment, job quality, and characteristics in 2011 and 2015. We can thus carry out a difference-in-differences analysis combined with a matching model to evaluate the impact of various types of innovation on employment and job quality. Given our methodological framework, our analysis is limited to the short-term effects of innovation. Nevertheless, we provide an alternative specification in the robustness checks that controls for repeated innovation over a longer period.

Empirical Strategy: A Difference-in-Differences Matching Model

Because firms that innovate have characteristics that differ from firms that do not, a simple variation of firms’ employment or job quality outcomes does not correspond to the proper effect of innovation: Employment or job quality trends are also influenced by many other factors, of which some may be observed (firm size, technological level, and so on), and others remain unobservable in the data. In addition, we also have to disentangle between the effect innovation has on job quality and the effect job quality may exert on innovation. For instance, workforce qualification level may favor firms’ innovations.

We use an empirical method that accounts for observable differences among firms (through propensity score matching) and corrects for unobserved characteristics (through difference-in-differences). The aim is to better approximate a causal effect of innovation, that is, the effect innovation has on employment and job quality, independently of firms’ characteristics (whether they are observable or not). This method proceeds in two steps. First, we use the firms’ characteristics, such as sector, size, age, and level of technology to predict whether firms will introduce innovations and to identify the characteristics of innovating firms. Among firms with these characteristics, some innovated and some did not innovate over the period considered for our empirical analysis. In the second step, we match firms that innovated with their “twins” that did not innovate and compare changes in job quality between the two groups of firms (that did or did not introduce innovations).

In technical terms, this method combines difference-in-differences with propensity score matching (PSM). To compare job-quality outcomes in similar firms, we use a PSM model initially developed by Rosenbaum and Rubin (1983) to assess the effects of medical treatments. This approach consists of considering innovation (I) as a treatment and constructing, for each firm that innovated between 2012 and 2014, a similar counterfactual firm that did not innovate.

The effect of innovation is measured by the outcome variable (here, differing measures of employment and job quality). Each firm thus has two potential outcomes: y 0 if I=0 and y 1 if I=1. The effect of innovation on job quality at the firm level is never observed, however, as y 0 and y 1 are never seen simultaneously; we only see the actual outcome.

Let Y i be the vector of result variables. For each firm, only the couple (Y, I) is observed. Nevertheless, if the latent outcome variables are independent of assignation to the treatment ( ( y 0 , y 1 ) ⊥ I ) —that is, if the treatment is randomly assigned—then the average effect on the treated firms (i.e., innovative firms) can be identified: E [ ( y 0 , y 1 ) / I = 1 ] . However, this independence property is unlikely to hold. One solution is to construct a control group so that the distribution of a set of observable characteristics (a set of control variables, denoted by X) is identical to that of innovating firms. This helps us to reduce the selection bias, as the identification condition becomes less restrictive, and the independence property should be checked ( ( y 0 , y 1 ) ⊥ I / X ). When many control variables are present, it can be difficult to find a counterfactual for each treated firm. According to Rosenbaum and Rubin (1983), conditional independence to the set of control variables implies independence relative to the propensity score, P ( X ) , which is a one-dimensional summary of the matching variables that estimates the probability of being assigned to the treatment conditional on these variables: ( y 0 , y 1 ) ⊥ I / X . To be valid, this method requires there be observations in both groups with similar propensity scores (a common support, which makes matching feasible), and that similar propensity scores are based on similar observed characteristics and covariates.

In practice, many propensity score matching methods are described in the literature. For instance, Caliendo and Kopeinig (2005) recommended using a number of estimators. We here use radius matching with a caliper of 0.00001, which is small and implies a precise matching between the treated and control firms. Indeed, as mentioned earlier, one of our main contributions is to capture as much as possible a causal effect of innovation on job quality, which requires comparing very similar firms. There is, however, a trade-off between the size of the matching group and the reduction of the bias between the treated and untreated firms. A small caliper value implies that more firms drop out of the common support (i.e., “off-support” firms) as they are too particular for counterfactuals to be found. In our case, the share of innovating firms dropping out of the common support is between 12.6% and 18.8% depending on the innovation variable we analyze, which we consider reasonable. Figure A.1 (in the Online Appendix) depicts the share of treated, untreated, and off-support firms by the predicted propensity score. Off-support firms do indeed have high propensity scores but no comparison firms. We test various methods and parameters in the robustness checks (kernel matching and radius matching with different calipers; see Discussion and Robustness Checks section).

We evaluate the matching robustness with a balancing test (see Online Appendix Table A.1) that analyzes the standardized differences. This method compares the mean of the control variables for the treated and untreated firms, and thus the reduction in the selection bias before and after matching. The results show that after matching, there are no average differences in the control variables between the treated (innovating firms) and control (non-innovating firms) groups. The selection model then reduces the bias between treated and untreated firms. Our choice of a small caliper value explains the considerable similarity between the two groups of firms, although it does exclude some very particular firms. Table A.2 in the Online Appendix reports the characteristics of off-support firms compared to on-support firms in the case of product innovation: ⁶ Off-support firms appear larger and belong more often to the high-technology sector. Choosing a very small caliper leads to dropping these big innovating firms and to focusing only on comparable firms to measure as much as possible a causal effect. This approach may, of course, change some of the effects usually found when big innovating firms are kept in the sample, but it gives a better measure of the proper effect of innovation.

A final condition for PSM validity is that no systematic differences should be between the treated and control groups in terms of unobserved characteristics that may influence the outcomes. This hypothesis may well not hold, as there are likely important unobserved factors influencing innovation behavior at the firm level: We therefore introduce difference-in-differences using the time dimension of our data to correct for unobserved heterogeneity. This consists of calculating the change in the outcome variable between two dates (the first difference) and comparing this change between the treated and untreated firms (the second difference). The formula for the treatment effect on the treated firms is as follows:

Δ = 1 N 1 ∑ i ∈ I 1 { ( Y t , i − Y t ′ , i ) − ∑ j ∈ I 0 M i [ [ ( P ( X j ) − P ( X i ) ) ] ∑ i ∈ I 0 [ ( P ( X j ) − P ( X i ) ) ] ] ( Y t , j − Y t ′ , j ) }

where N 1 is the number of firms that innovate; I 1 , the sample of innovating firms; I 0 , those that do not innovate; P(X), the estimated propensity score; and Y, the outcome variable of employment or job quality. Last, M ⁱ [] is the average value of the outcome variable for the population of firms j that belong to the control group selected from firms i, and t and t′ are the periods before (2011) and after (2015) treatment, respectively. This estimator is intended to satisfy the common-trend assumption so that the trend in the outcome variables before innovation is the same in the treated and control firms. We check this assumption in our data for employment, hourly wages, and working hours from 2006 to 2015. ⁷ The common trend holds for employment, working hours and the hourly wage (see Figures A.2 and A.3 in the Online Appendix for product and process innovation). Testing this assumption is not obvious in our case, however, as innovation can be repeated and the two groups of firms may have already innovated before our observation period. To overcome this limitation, we also run regressions on a more limited subsample of firms from 2009 to 2015 for which we have information about repeated innovation (see Discussion and Robustness Checks section).

Our empirical strategy consists of two steps. In the first, we estimate a logit model to produce the propensity score. We here consider various determinants of innovation: firm size and age, industry, whether the firm is part of a business group, labor costs, and labor productivity. All of these variables are measured in 2011, that is, before the firms decide whether to innovate. For age, productivity, and labor costs, we calculate quartiles over the sample as explanatory variables. In the second step, we estimate the average effect of the treatment—the treatment here being innovation—on the difference in employment and job-quality changes for the treated and control groups using the radius-matching estimator.

Descriptive Statistics

In our database, we identify the firms that innovated between 2012 and 2014 (Table 1): 27.6% of the firms declared to have introduced a new or significantly improved product, and 27.5% introduced a new or significantly improved process. ⁸ A total of 18.9% of the firms developed product innovations that they declared to be “new to the market,” which we consider as an indicator of the novelty or intensity of innovation. Far fewer firms (7.3%) declared both product innovation and patent application, which also indicates more intensive innovation activity.

OPEN IN VIEWER

Table 1.

Share of Innovating Firms by Type of Innovation and across Industries (between 2012 and 2014)

Type of innovation	Total (%)	Manufacturing (%)	Services (%)
Product innovation	27.6	42.1	26.2
Product innovation–new-to-the-market	18.9	31.0	16.8
Product and patenting firms	7.3	16.8	3.6
Process innovation	27.5	40.0	24.7

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 14,491 firms.

The innovating firms have particular characteristics. They are overrepresented in manufacturing but underrepresented in retail and construction. We found that 47.0% of product innovators and 44.9% of process innovators are in manufacturing, and this figure is higher for the more intensive types of innovation (50.5% of new-to-the-market product innovators, 70.8% of product innovation and patenting firms). The share of innovating firms in services is slightly below their sample share, but the difference is small. Decomposing by technology, the innovating firms are overrepresented in all groups except less knowledge-intensive services. Larger firms (more than 50 employees) are overrepresented among innovators (57.1% of product innovators and 53.4% of process innovators have more than 50 employees, compared with 37% in the whole sample), as well as members of a business group (65.2% of product innovators and 60.9% of process innovators). Innovating firms are older on average and have higher average labor costs and productivity. ⁹

We can use our matched sample of firms to compare a number of indicators of employment and job quality in innovating and non-innovating firms. Some of the main indicators for 2011 are summarized in Table 2.

OPEN IN VIEWER

Table 2.

Job Quality and Employment by Firm Innovation Status

	Total	Product innovation		Process innovation		Product innovation–new-to-the-market		Product innovation and patent
		No	Yes	No	Yes	No	Yes	No	Yes
Total workforce (number of employees)	235.1	152.0	453.1	160.2	432.6	166.2	531.4	188.9	821.6
Permanent contracts (% share)	94	94	95	94	95	94	96	94	97
Temporary contracts (% share)	6	6	5	6	5	6	4	6	3
Managers and professionals (% share)	17	13	25	15	22	14	28	16	29
Technicians and associate professionals (% share)	17	16	21	16	20	16	22	17	24
Manual and clerical workers (% share)	65	70	53	68	58	69	49	67	47
Number of hours (annual)	1,826.7	1,825.6	1,829.8	1,827.1	1,825.9	1,826.0	1,830.0	1,826.2	1,834.0
Hourly gross wage (in euros)	19.2	18.3	21.7	18.6	20.8	18.5	22.6	18.8	24.6

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 14,491 firms.

In Table 2, the shares of open-ended and temporary contracts are similar in the subsamples of innovating and non-innovating firms, for product and process innovation. In the case of more intensive product innovation (new-to-the-market or patent), the share of permanent contracts appears higher in innovating firms. Average annual hours worked per employee are close in innovating and non-innovating firms. Hourly wages are systematically higher in innovating firms, with the gap being higher for more intensive innovators (product–new-to-the-market, and especially product and patenting). Innovating firms also have workforce skill structures that differ from those of non-innovating firms: ¹⁰ They have smaller shares of manual and clerical workers but more managers and professionals as well as more technicians and associate professionals.

The Determinants of Innovation at the Firm Level

The logit regressions (for the four innovation types defined above) include some structural firm characteristics that are correlated with innovation, such as industry decomposed by level of technology, firm size and age, and whether the firm belongs to a business group. Productivity is also introduced as an economic performance indicator, as well as labor costs that are usually considered as a factor determining innovation capacity. We introduce the corresponding quartiles for the continuous variables (age, productivity, and labor costs). Table 3 shows the results for product innovation (the results for the other innovation variables appear in Table A.3 in the Online Appendix).

OPEN IN VIEWER

Table 3.

Determinants of Firm Product Innovation

Dependent variable	Product innovation
Size (ref. 10 to 19 employees)
20 to 49	0.25 (0.06)***
50 to 499	0.93 (0.06)***
500 to 999	1.35 (0.09)***
>1000	1.82 (0.10)***
Member of a business group (ref. No)
Yes	0.33 (0.05)***
Sector by level of technology (ref. Less knowledge-intensive services)
High-tech (Manufacturing)	2.04 (0.15)***
Medium high-tech (Manufacturing)	1.79 (0.08)***
Medium low-tech (Manufacturing)	0.82 (0.07)***
Low-tech (Manufacturing)	0.44 (0.06)***
Knowledge-intensive services	0.79 (0.06)***
Age (ref. lowest quartile)
2nd quartile	−0.08 (0.06)
3rd quartile	−0.12 (0.06)**
Top quartile	−0.12 (0.06)**
Productivity (ref. lowest quartile)
2nd quartile	0.05 (0.07)
3rd quartile	0.15 (0.07)**
Top quartile	0.30 (0.08)***
Labor cost (ref. lowest quartile)
2nd quartile	0.09 (0.07)
3rd quartile	0.22 (0.08)***
Top quartile	0.38 (0.09)***
Intercept	−2.45 (0.07)***
Number of observations	14 491
LR χ2(18)	2369.02
Prob > χ2	0.0000
Pseudo R2	0.1388
Log likelihood	−7350.23

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 14,491 firms.

***

p < 0.01; **p < 0.05; *p < 0.1.

As in the descriptive analysis, all types of innovations occur more in larger firms, which can be explained by large fixed innovation costs and there being more employees dedicated to innovative work. Being in a group also increases the probability of innovation. Compared to less knowledge-intensive services (the reference category), innovation is more likely in all other industries. Innovation increases with the technological level and is higher in manufacturing (compared to services). The effect of age differs from that in the descriptive statistics: Once we control for other firm characteristics, older firms are less innovative. This outcome may reflect a Schumpeterian effect in that new firms compete with older firms by introducing new products or processes. High productivity (the fourth quartile of ex ante 2011 productivity) increases product innovation. Higher labor costs also increase innovation. In a given institutional context (French firms generally face the same law in terms of minimum wage and social contributions ¹¹ ), higher labor costs are actually a sign of a more educated workforce, which drives more innovation from a human capital perspective.

These results for the factors explaining innovation are very similar to those for new-to-the-market product innovation and process innovation (see Online Appendix Table A.3). ¹²

Fairly Positive Impact of Innovation but Strong Differences by Occupation

The second step compares the 2011 to 2015 changes in employment and job quality between innovating (treated) and non-innovating (control) firms. We also decompose the employment and job-quality effects (and in particular, wages) of innovation by occupation.

We first find a positive and significant impact of product innovation on firm employment, which also appears when firms have applied for a patent (see Table 4). More precisely, that positive impact corresponds to a positive difference in employment variations between innovating and non-innovating firms. In the case of process innovation, the impact on employment is negative (i.e., the difference in employment variations between innovating and non-innovating firms is negative). These results appear to be consistent with the existing theoretical and empirical literature that generally finds a more positive effect of product innovation at the firm level, whereas the results are more mixed for process innovation. The average size of the results is rather small but grows with the intensity of innovation: The workforce increases amount to 5.7 employees for product innovation and 14.4 employees for product innovation in patenting firms. ¹³

OPEN IN VIEWER

Table 4.

Impact of Innovation on Employment and Job Quality

Dependent variables	Product	Process	Product–new-to-the-market	Product-patenting firms
Total workforce	5.7 (3.0)*	−5.6 (2.8)**	3.2 (4.0)	14.4 (8.1)*
Open-ended (permanent) contract employees	8.2 (3.0)***	−2.9 (2.7)	5.7 (4.0)	19.8 (7.8)**
Fixed-term contract employees	−1.7 (1.1)	−3.0 (1.1)***	−1.1 (1.2)	−2.0 (2.0)
Average annual hours worked per employee	12.7 (6.8)*	0.4 (7.0)	8.6 (7.3)	10.9 (10.0)
Hourly wages (gross)	−0.0 (0.1)	0.0 (0.1)	0.0 (0.1)	0.2 (0.1)

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 14,491 firms.

Notes: The first three variables represent the variations in the number of employees. The fourth variable shows the difference in hours (per employee) and the last one represents the difference in euros (per employee). These results are from difference-in-differences models, psmatch 2.

***

p < 0.01; **p < 0.05; *p < 0.1.

When decomposing employment by type of labor contract (permanent or temporary), we observe a positive effect of innovation on permanent-contract employees both for product innovation and combined with patent application, whereas that for temporary-contract employees is either insignificant (for product innovation) or negative (for process innovation). Concerning permanent contracts, the effects are stronger than in the case of total employment: Product innovations increase the number of permanent employees by 8.2 and 19.8 in the case of patenting firms. The negative effect of process innovation on fixed-term contracts is more limited (–3 employees). Technological innovation (product as well as process) then favors stable employment (at least in terms of labor contracts). Innovative firms appear to invest in their human capital rather than increase labor flexibility.

In terms of working hours, we report a positive impact of total product innovation on annual hours, which remains relatively small (+12.7 hours on average, to be compared with the 1,826.7 hours worked annually in sample firms). The effect is insignificant for process innovation as well as for more intensive forms of product innovation. This positive effect for product innovation may reflect a number of phenomena that we cannot disentangle: an increase in the share of full-time workers or more hours being worked by part-time and/or full-time workers.

Last, the impact of product and process innovation on hourly wages is insignificant: In our sample, innovation activities thus do not seem to produce any rent-sharing with workers. This insignificant result is not surprising to us since theoretical literature indicates ambiguous effects of innovation on wages (Calvino and Virgillito 2018). As mentioned previously, an increase in wages can appear if workers are able to appropriate some of the productivity gains. It will thus depend on their bargaining power. In our case, the relatively short time horizon of our study does not necessarily enable workers to exert their bargaining power and negotiate higher wages, especially in the aftermath of the 2008 financial crisis. In addition, an increase in the workforce may hide some flows in and out of the firm so that we cannot measure the impact on the wages of employees who are staying (and should be able to capture part of the productivity gains).

To sum up, the empirical results underline ambiguous effects of technological innovation on employment, somehow positive in the case of product innovation but negative for process innovation. They also show a fairly positive impact of both product and process innovation on the stability of employment contracts, which are an important component of job quality. No effect on wages is found and a limited positive effect on working hours is found, however, so job quality effects appear limited (although rather positive). In the case of product innovation only, these results tend to confirm the hypothesis that innovation could produce a virtuous circle at the firm level, favoring both the quantity and quality (especially here, stability) of jobs. This interpretation is, of course, speculative, as the definition of job quality is restricted to only a few dimensions and does not include important components such as working conditions. We, in addition, do not know whether temporary contracts and low working hours are voluntary or involuntary. Nevertheless, given the importance of job security for French workers and the very protective legislation in France regarding working hours, these features do seem central to worker well-being.

The overall employment and job-quality effects of innovation may also differ according to skill level. Apart from contract type, we can decompose these effects by occupation.

For employment (Table 5), notable differences by occupational group are evident. First, all types of product innovation increase the number of managers and professionals as well as intermediate occupations, ¹⁴ compared to non-innovating firms, whereas this is not the case for manual and clerical workers (the effect is insignificant). Second, process innovation significantly decreases the number of manual and clerical workers, whereas its effects on other occupational categories are insignificant. Thus, a skill-biased pattern occurs in the employment effects of firm-level innovation: Product and process innovations are associated with skill upgrading and benefit higher-skilled workers. According to the estimations, the growth in the number of managers and professionals is stronger for patenting firms (+15 managers and professionals compared to +5.6 for total product innovation). Given the positive impact on the employment of intermediate occupations obtained for product innovation (in general or combined with patent application), our results do not correspond to the hypothesis of skill polarization from innovation but are rather in line with skill-biased technological change.

OPEN IN VIEWER

Table 5.

Impact of Innovation on Employment by Occupations

Dependent variables	Product	Process	Product–new-to-the-market	Product-patenting firms
Number of managers and professionals	5.6 (1.6)***	0.4 (1.6)	6.5 (2.5)***	15.0 (5.9)**
Number of intermediate occupations	2.8 (1.1)**	−0.2 (1.0)	1.3 (1.5)	6.6 (3.3)**
Number of manual and clerical workers	−2.8 (2.9)	−5.8 (2.7)**	−4.5 (3.8)	−7.1 (8.4)

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 14,491 firms.

Notes: These results are from difference-in-differences models, psmatch 2.

***

p < 0.01; **p < 0.05; *p < 0.1.

Regarding wages and working time, most effects by occupational group are insignificant, but product innovation leads to slightly lower wages ¹⁵ for manual and clerical workers (see Table A.4 in the Online Appendix). This can also be read as skill bias in the effects of innovation that reduces wages for less-qualified workers. This effect is only found for total product innovation though and disappears for the more intensive forms of innovation (new-to-the-market and patenting firms). Working time also increases for lower-skilled occupations following product innovation, indicating that firms tend to prefer longer hours for these categories of workers rather than hiring new employees or increasing wages.

Heterogeneity in Job-Quality Outcomes by Industry

Because the general results presented above may conceal considerable industry-level heterogeneity, we re-estimate our innovation regressions by industry. ¹⁶ The results appear in Tables 6 and 7.

OPEN IN VIEWER

Table 6.

Impact of Innovation on Employment and Job Quality: Manufacturing

Dependent variables	Product	Process	Product–new-to-the-market	Product-patenting firms
Total workforce	15.7 (3.2)***	−1.8 (3.1)	9.4 (4.5)**	22.7 (7.6)***
Open-ended (permanent) contract employees	16.9 (3.2)***	−5.2 (3.0)	11.3 (4.4)**	25.1 (7.4)***
Fixed-term contract employees	−2.8 (0.9)***	−2.0 (0.9)**	−4.0 (1.2)***	−3.9 (2.1)*
Average annual hours worked per employee	25.6 (10.0)**	−5.3 (8.6)	16.2 (9.9)	5.0 (11.3)
Hourly wage (gross)	−0.2 (0.09)*	0.1 (0.1)	−0.1 (0.1)	0.1 (0.1)

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 5,058 firms.

Notes: The first three variables represent the variations in the number of employees. The fourth variable shows the difference in hours (per employee) and the last one represents the difference in euros (per employee). These results are from difference-in-differences models, psmatch 2.

***

p < 0.01; **p < 0.05; *p < 0.1.

OPEN IN VIEWER

Table 7.

Impact of Innovation on Employment and Job Quality: Services

Dependent variables	Product	Process	Product–new-to-the-market	Product-patenting firms
Total workforce	−1.3 (5.4)	−15.4 (5.5)***	−4.1 (7.8)	4.6 (23.4)
Open-ended (permanent) contract employees	2.4 (5.4)	−8.4 (5.4)	−0.4 (8.1)	18.0 (22.4)
Fixed-term contract employees	0.1 (2.5)	−5.2 (2.7)*	4.0 (2.2)*	4.0 (3.4)
Average annual hours worked per employee	8.7 (12.3)	0.6 (15.9)	4.7 (13.4)	13.0 (21.4)
Hourly wage (gross)	−0.0 (0.1)	−0.04 (0.1)	0.1 (0.1)	0.6 (0.4)*

Sources: CIS 2014; FARE 2011–2015; and DADS 2011–2015. Matched data based on authors’ calculations for 4,462 firms.

Notes: The first three variables represent the variations in the number of employees. The fourth variable shows the difference in hours (per employee) and the last one represents the difference in euros (per employee). These results are from difference-in-differences models, psmatch 2.

***

p < 0.01; **p < 0.05; *p < 0.1.

The results for manufacturing are similar to those in the whole sample but display stronger and more significant effects on employment. Product innovation increases total and permanent employment but reduces fixed-term employment. The positive effects on employment are larger than those observed for the whole sample (+15.7 employees and +16.9 permanent contracts in the case of product innovation, +22.7 employees and +25.1 permanent contracts in the case of product innovation in patenting firms). The negative impact of process innovation on total employment becomes nonsignificant when focusing on manufacturing, but the negative effect on fixed-term contracts is confirmed. Only a few significant estimated coefficients appear in the service sector, and these confirm general results for process innovation (negative effect on total employment, which appears stronger than for the whole sample, and negative effect on fixed-term contracts), but often differ from those in manufacturing: In particular, the new-to-the-market product innovation positively affects fixed-term contracts.

The results for other job quality variables are less clear-cut. Considering working hours, the only significant effect is an increase in the case of total product innovation in manufacturing (as in the general results). As far as wages are concerned, a slightly negative effect on wages appears for product innovation in manufacturing (only in the general case, not for the more intensive forms of innovation), whereas for services, we find a positive effect for product innovation in patenting firms. Although both effects are very small (–0.2 euros per hour for the first, and +0.6 euros for the second), they suggest that wage effects of innovation are not homogeneous across industries.

Although the literature on innovation effects by industry is limited, the differences between manufacturing and services may be interpreted with regard to some hypotheses mentioned before. In particular, the insignificant effect of product innovation in services may be related to lower competition compared to manufacturing. However, services remain a very broad category and this result could also hide varying effects in subcategories of services. When we run regressions only in high-tech or low-tech services, ¹⁷ it appears that the increase in fixed-term contracts and the slight increase in wages is concentrated in low-tech services, whereas in high-tech services, the results are slightly closer to manufacturing (with a positive effect of product innovation on employment). These results on low- and high-tech services are in line with the rare previous studies that analyzed the effect of innovation distinguishing by technology level of services using CIS (Evangelista and Savona 2002). These results also show that the labor-saving hypothesis on process innovation seems to hold for services and is concentrated in low-tech services.

Comparing the results between manufacturing and services suggests contrasting human resources strategies coexist regarding innovation. In manufacturing, firms increase the workforce (number of employees and working hours) and invest in employment stability, whereas in services (especially in low-tech services), they develop rather flexible jobs and slightly increase wages. Wages and employment flexibility can be related both directly and indirectly. In accordance with French labor law, fixed-term contracts come with a flexible wage premium, and more generally, wages may be a way of compensating for greater job flexibility.