The Importance of Process and Product Innovation for Productivity in French Manufacturing and Service Industries
The Importance of Process and Product Innovation for Productivity in French Manufacturing and Service Industries
Abstract and Keywords
This chapter examines the effect of innovation on labour productivity in France, using a general framework that accounts for research activities and for both product and process innovation. To control for selection and endogeneity effects, the chapter estimates a nonlinear multipleequations econometric model. This model is estimated on the third and fourth waves of the French component of the Community Innovation Survey, using a threestep and a twostep estimation procedure. The chapter identifies process innovation as the main driver of labour productivity in the manufacturing industry, but the results are less consistent in the services industry. Sensitivity analyses actually suggest that our two indicators of innovations both tend simply to measure overall innovation. When a single indicator is used, the chapter finds a positive effect of innovation in all industries and in all periods. The chapter concludes with some cautionary words relating to the joint modelling of the impacts of product and process innovation on productivity.
Keywords: R&D, product innovation, process innovation, knowledge production function, innovation and productivity
6.1. Introduction
This chapter investigates the effect of innovation on the productivity of firms when two technological paths are available: product innovation on the one hand, and process innovation on the other. This is done by estimating a nonlinear multipleequation econometric model, which allows us to control for both selection and endogeneity. This model encompasses three stages. The first stage deals with R&D activities, in terms of both propensity and intensity. The second stage represents the ‘knowledge production function’, distinguishing between product and process innovation. The third and final stage identifies the impact of both types of innovation on labour productivity.
This model is estimated on two waves of the French component of the Community Innovation Survey (CIS): CIS3, which covers the years 1998–2002, and CIS4, which covers the years 2002–2004. Using both waves of the survey, we conduct an intertemporal comparison on the manufacturing industry. Using the fourth wave, we also compare the innovation–productivity relationship in the manufacturing industry and in the services. Process innovation appears as the main driver of labour productivity in the manufacturing industry in both periods, but further analyses reveal that it is actually difficult to disentangle the respective effects of product and process innovation. Both types of innovation actually seem to capture ‘overall’ innovation, and when a single indicator of innovation is used, it always has a (p.129) significantly positive impact on labour productivity, both in the manufacturing industry and in the services.
The remainder of this study is organized as follows. In Section 6.2, after discussing issues pertaining to product and process innovation in our general framework of the R&D–innovation–productivity relationship, we detail the specification of the econometric model, and present the estimation procedures. In Section 6.3, we describe the data and define the explanatory variables used in the various equations of the model. Section 6.4 is dedicated to the presentation and interpretation of the results, and to some sensitivity analyses. We conclude in a final section.
6.2. Motivation and econometric modelling
6.2.1. Accounting for product and process innovation in the CDM model
Our contribution brings together two strands of literature: (1) empirical studies of the relationship between R&D and the output of innovation, and (2) studies dealing with the impact of innovation on productivity. To do so, we follow the approach initially proposed by Crépon et al. (1998), who developed a ‘structural’ model linking R&D, innovation and productivity. This model—usually referred to as the ‘CDM’ model—has enjoyed significant success in the literature over the recent years. CDMtype models are generally built as threestage econometric models that relate productivity to new knowledge that depends on firms’ R&D effort, which is in turn determined by a number of firm and environmentspecific factors.
Recent applications of this framework include Griffith et al. (2006), who attempt to extend this framework by taking into account process as well as product innovation. They estimate their variant of the CDM model in four EU countries (France, Germany, Spain, and the United Kingdom). Chudnosky et al. (2006) estimate a similar model on Argentine firmlevel data. In addition, they propose a short summary of CDMtype studies implemented between 1998 and 2006. For an overall survey, see also Mairesse and Mohnen (2010). Other recent studies based on a CDM approach include Mairesse et al. (2005),Raffo et al. (2008), and Hall et al. (2009).
Our variant of the CDM model follows in the footsteps of Griffith et al. (2006), and tries for a number of improvements in both the specification and the estimation methods. Before detailing those, it may be useful to discuss the importance of adequately taking into account both product and process innovation in a CDMtype model.
(p.130) While economic theory offers little guidance regarding the respective (and potentially different) impacts of product and process innovation on firms’ productivity, empirical applications such as ours face a recurrent measurement problem. When a productivity measure is computed using firmlevel data, this measure is generally based on a value expressed in monetary units (such as total sales or value added). Information on individual prices is generally not available in microeconomic data (especially in survey data), making it impossible to measure firms’ output in real terms of say ‘quantity’ or ‘volume’, of goods produced (see Griliches, 1979; Mairesse and Jaumandreu, 2005). The distinction between volume and valuebased measures of productivity has important implications for empirical studies of the impact of innovation, all the more so in studies which distinguish between product and process innovation.
Indeed, while both types of innovation could, in theory, lead to productivity gains that are difficult to assess in empirical studies relying on valuebased measures of productivity. First of all, a process innovation reduces the costs of production, which can lead a firm to lower its output price. Since we do not measure the firm’s output in real terms, but only its global sales or value added, we may observe a negative impact of process innovation on productivity simply because the firm is now selling the same output at a lower price. This negative impact is, however, artificial, and only arises because we are not able to disentangle firm sales into an output level and a price. In this situation, process innovation would actually have improved the firm’s competitiveness, even though we would observe a negative impact of process innovation on our (imperfect) measure of productivity.
The same measurement issue can lead to the reverse problem as far as product innovation is concerned. By definition, a product innovation leads to a ‘new’ or ‘improved’ product, which is considered as ‘better’ either because it presents a new feature and its quality has actually been improved, or because it simply looks different. A firm may decide to sell this product at a higher price to reflect the fact that it is a new, a better quality or simply a different product. In this case, we may observe an artificially high effect of product innovation on productivity, again because we are not able to disentangle prices and output levels in our valuebased measure of productivity. This phenomenon may be even more important in the services industry, where prices are often supposed to signal the quality of the services that are delivered.
The aforementioned difficulties can be aggravated by the fact that product and process innovations are often concurrent, because the introduction of a new product often requires some significant change in the process of production. This can lead to a correlation between indicators of product and process innovation, which may explain why empirical studies often diverge in their conclusions regarding their respective impacts on productivity. A similar analysis can even lead to different conclusions across different samples. For (p.131) instance, in their version of the CDM model, Griffith et al. (2006) find a positive impact of process innovation on labour productivity in France, but no impact whatsoever in Germany, Spain, or the UK. Raffo et al. (2008), who estimate a similar CDMtype model, find a positive impact of process innovation in France, Spain, Brazil, and Mexico, but no impact in Switzerland or Argentina. Although we are not able in the present study to avoid the abovementioned difficulties, we must be aware of them in the interpretation of our results.
6.2.2. Econometric specification
Our aim is to estimate a variant of the CDM model accounting for both product and process innovation, in the line of Griffith et al. (2006). Our model can be written as a recursive system of five econometric equations,^{1} three of which are nonlinear:
where 1(.) denotes the indicator function, which is equal to 1 if its argument is true, and to 0 otherwise. The x_{ki}’s (with k = 1,…, 5) are vectors of explanatory variables, and the u_{ik}’s (with k = 1,…, 5) are randomerror terms. Finally, β_{k} (with k = 1,…, 5) denotes a vector of parameters to be estimated, while α_{3}, α_{4}, γ, and δ are single parameters to be estimated. We relied on different procedures to estimate System (1). Before detailing these estimation procedures, let us examine what each of the five equations represents.
The first equation in System (1) accounts for selection into R&D. It explains the probability that firm i does R&D on a continuous basis, and is specified as a probit model, i.e. P(r_{i} = 1) = Φ(x _{1i} β _{1}), where r_{i} = 1 if firm i reports R&D expenditures, and r_{i} = 0 otherwise. The second equation describes the R&D effort of firm i, conditional on doing R&D. It is specified as a linear equation, which relates the log of R&D intensity (defined as the ratio of R&D expenditures to the number of employees) to a number of potential determinants x _{2i}. Taken together, and assuming that u _{1} and u _{2} are bivariate normal with (p.132) correlation coefficient ρ _{12}, these two equations define a generalized (or type II) tobit model.
The third equation in System (1) models the probability that a firm introduces a product innovation over a period of time (equal to three years in the CIS). The fourth does the same with regard to process innovation. Each one of these two equations is specified as a probit model which includes an endogenous regressor and several exogenous regressors. The endogenous regressor is ln rd_{i}, the log of R&D intensity. The exogenous regressors are control variables included in the x_{3i} and x_{4i} vectors respectively. Taken together, and assuming that u _{3} and u _{4} are correlated with correlation coefficient ρ _{34}, these two equations define a bivariate probit model. This specification takes into account the fact that product and process innovation can be jointly determined.
Economically, this bivariate probit model represents what is known in the literature as the ‘knowledge production function’. The knowledge production function relates measures of new knowledge (innovations) to innovation inputs (such as R&D expenditures). In our representation of the knowledge production function, ln rd_{i} is potentially endogenous. Indeed, unobservable characteristics could increase both a firm’s R&D effort and its ‘innovativeness’ or ‘innovativity’ (i.e. its ‘productivity’ in producing innovations, as defined in Mairesse and Mohnen, 2002; and Mohnen et al., 2006). We explain below how each estimation procedure deals with this potential endogeneity.
The fifth and final equation in System (1) models the log of labour productivity LP_{i} (defined as the ratio of value added to the number of employees) as a function of product innovation, process innovation and a number of control variables (including a proxy for physical capital). Again, product (and process) innovation is potentially endogenous as unobserved factors could influence both the innovation process and the production process. We explain below how each estimation procedure deals with this potential endogeneity. Economically, the fifth equation in System (1) derives from a CobbDouglas type production function where the main inputs are labour, capital, and knowledge (represented by the two indicator variables prod_{i} and proc_{i}). The role of materials as an input is implicitly taken into account through the use of value added (rather than turnover) to define labour productivity.
6.2.3. Estimation procedures
The original CDM model was estimated simultaneously using asymptotic least squares (ALS). Although system estimators provide a gain in efficiency, most subsequent applications of the CDM framework (such as the ones mentioned in Section 6.2.1) usually estimate the three stages sequentially, with the predicted output of a given stage being used as an explanatory variable in the next (p.133) one. In the present research, we adopt two distinct procedures to estimate this model: (1) a threestep sequential estimation procedure similar to that used by Griffith et al. (2006), and (2) a twostep sequential estimation procedure in which the innovation production function and the productivity equation are jointly estimated. The former is supposedly more robust but less efficient, whereas the latter is potentially more efficient but less robust. In our empirical application, we will be able to check whether these two procedures lead to consistent results.
The threestep sequential estimation procedure is the most commonly used in the literature. In the first step, we estimate the generalized tobit model which describes both selection into R&D and R&D intensity. In the second step, we estimate the bivariate probit model which depicts the knowledge production function, using the predicted (rather than actual) value of R&D intensity as a regressor.^{2} The third and final step consists of estimating the labour productivity equation, using the predicted values of the probabilities to innovate as regressors (rather than the indicators). We use maximumlikelihood estimation in the first two steps, and ordinary least squares (OLS) in the third one. In this estimation procedure, we address the abovementioned endogeneity issues in the second and third steps by replacing the potentially endogenous regressors with their predicted values (obtained from the previous step). Standard errors are bootstrapped to correct for the bias induced by the inclusion of predicted regressors. For identification, we rely on ‘instruments’ included in the previous step and excluded from the current step (the list of instruments for each step will be detailed in Section 6.3.2).
The twostep sequential estimation procedure is a potentially more efficient variant of the threestep procedure described above. The first step is actually identical in both procedures: we estimate a generalized tobit model to describe both selection into R&D and R&D intensity. In the twostep procedure, however, the knowledge production function and the productivity equation are jointly estimated by maximum likelihood, using the predicted value of R&D intensity in the product and processinnovation equations. In other words, the second step of this procedure consists in estimating by maximum likelihood the following threeequation model, which combines two nonlinear (probit) equations and a linear equation:
where ln r^ _{i} denotes the predicted value of ln rd_{i} (obtained from the first step of the procedure). In Model (2), u_{i} _{3}, u_{i} _{4}, and u_{i} _{5}, the error terms, are assumed to be trivariate normal. This implies that the contributions to the likelihood will be connected by the correlation coefficients of the error terms. There are three components to the likelihood, two of which correspond to the likelihood of a probit and one to the likelihood of a linear model. Thus, the likelihood of Model (2) can be written as:
where ρ_{jk} denotes the correlation coefficient between errors u_{j} and u_{k} (with j = 3, 4, 5; and k = 3, 4, 5), and σ _{5} the standard deviation of random error u _{5}. By convention, ϕ and Φ respectively denote the density and c.d.f. of the normal distribution.
In practice, Model (2) is simultaneously estimated by fullinformation maximum likelihood, using the conditional mixed process application (CMP) developed by Roodman (2009). Maximizing the loglikelihood of this model requires solving a multiple integral of dimension 3, which cannot be done analytically. To solve this problem, the CMP program relies on a GHKtype numerical simulation algorithm.
In theory, we could expect the twostep procedure to be somewhat more efficient, but potentially less robust, than the threestep procedure. From an empirical perspective, we are interested in checking whether the two procedures give similar results (in terms of both the signs and magnitudes of the estimated parameters of interests). The empirical application is conducted on the third and fourth waves of the French component of the CIS, in both the manufacturing (for CIS3 and CIS4) and services (for CIS4 only) industries. This will allow us to observe the evolution of the innovation–productivity relationship in the French manufacturing industry over two periods of time (1998–2000 and 2002–2004). We will also be able to compare this relationship in the manufacturing and services industries in the most recent observation period (2002–2004).
(p.135) 6.3. Data and variables
6.3.1. The French CIS3 and CIS4 databases
The present study uses firmlevel data from the third and fourth waves (CIS3 and CIS4) of the French component of the CIS. The CIS is a harmonized survey that is carried out by national statistical agencies in all EU member states under the coordination of Eurostat. CIS3 and CIS4 were conducted in 2001 and 2005 respectively and provide information for the periods 1998–2000 and 2002–2004. Both waves of the survey provide information on firms’ R&D activities, sources of knowledge, intellectualproperty protection, product and process innovations, other forms of innovation (e.g. organizational changes) and abandoned innovations.
However, the two waves present some key differences. The French CIS3 gives some information about firms’ investment in physical capital. It also allows researchers to distinguish between different types of human capital (i.e. lowskill and highskilled, the latter being measured by the proportion of employees with higher education in the workforce). The information about physical and human capital has disappeared in CIS4, which has been extended into two other directions: firstly, CIS4 samples firms with 10 employees or more, whereas CIS3 only included firms with 20 employees or more; secondly, CIS3 was focused mostly on manufacturing firms, whereas CIS4 covers the services industry quite extensively.
This structure of the data will lead us to use CIS3 and CIS4 to draw two different comparisons. First, we will examine how the relationship between R&D activities, innovation, and productivity in the French manufacturing industry has evolved across time, i.e. from the years 1998–2000 to the years 2002–2004. To do so, we will compare the results obtained with the sample of CIS3 manufacturing firms, to those obtained with the sample of CIS4 manufacturing firms. Second, we will use CIS4 to examine whether this relationship differs between the manufacturing and services industries in the recent period.
In order to be able to draw relevant conclusions from these comparisons, we selected three samples of firms with 20 employees of more from both CIS3 and CIS4: two manufacturing samples (a sample of 3524 firms from CIS3 and a sample of 4955 firms from CIS4) and a services sample (3599 firms from CIS4). Table 6.1 gives a breakdown of these three samples by category of industry. These industry categories are based on the twodigit NACE classification, aggregated in a fashion similar to that of Griffith et al. (2006). We cleaned all three samples in the usual way: we deleted a few observations with extreme values of turnover, observations where the absolute value of the rate of growth (p.136)
Table 6.1. Number and proportion of firms by industry in manufacturing and services for CIS3 and CIS4
CIS3 
CIS4 


Before matching 
After matching 
Before matching 
After matching 

Industry 
Nace 
N 
% 
N 
% 
N 
% 
N 
% 
Textile 
17–19 
475 
13.5 
475 
13.5 
648 
13.1 
636 
13.2 
Wood/Paper 
20–22 
386 
11.0 
386 
11.0 
706 
14.2 
654 
13.5 
Chemicals 
23–24 
374 
10.6 
374 
10.6 
451 
9.1 
445 
9.2 
Plastic/Rubber 
25 
276 
7.8 
276 
7.8 
313 
6.3 
311 
6.4 
Nonmetallic min. 
26 
168 
4.8 
168 
4.8 
273 
5.5 
265 
5.5 
Basic metals 
27–28 
607 
17.2 
607 
17.3 
699 
14.1 
684 
14.2 
Machinery 
29 
418 
11.9 
413 
11.7 
439 
8.9 
433 
9.0 
Electrical 
30–33 
452 
12.8 
452 
12.9 
657 
13.3 
644 
13.3 
Vehicles 
34–35 
192 
5.4 
191 
5.4 
483 
9.7 
475 
9.8 
Miscellaneous 
36–37 
176 
5.0 
176 
5.0 
286 
5.8 
284 
5.9 
All Manufacturing 
3,524 
100 
3,518 
100 
4,955 
100 
4,831 
100 

Hotels/Restaurants 
55 
505 
14.0 
482 
13.9 

Communication 
64 
78 
2.2 
72 
2.1 

Housing/Real Estate 
70 
269 
7.5 
258 
7.4 

Rental 
71 
179 
5.0 
171 
4.9 

ICT services 
72 
406 
11.3 
397 
11.5 

Services to firms 
74 
2,162 
60.1 
2,087 
60.2 

All Services 
3,599 
100 
3,467 
100 
Notes: The industry definition is based on the classification system NACE (Nomenclature générale des activités économiques dans les Communautés Européennes) as published by Eurostat, using twodigit levels. ‘Matching’ refers here to the matching of our survey data with administrative data on firms (which provided information on physical capital and value added).
Besides clean samples, our analysis also requires information on physical capital that is not available in CIS4. We obtained this information by matching CIS4 with administrative data from the French yearly firm census (EAE).^{3} In order to have similar variables in all samples, we also matched CIS3 with the same administrative data (over a different period). This provided us, for each wave of the CIS, with two possible proxies for physical capital, observed in t, t1, and t2 (t being the year the CIS was conducted). The first proxy is investment, and the second one is the book value of fixed assets. Matching CIS3 and CIS4 with the relevant yearly firm survey also provided us with value added, which we use to compute labour productivity. Using value added rather than turnover allows us to control implicitly for the effect of materials in the labour productivity equation of the model.
As can be seen in Table 6.1, matching the CIS data with administrative data did not cause a greater loss of observations that the loss which would have resulted from a simple cleaning of the data. In any sample, we lose very few firms from matching and the distribution of firms across industry categories remains the same after matching the data.^{4} This suggests that we do not have to fear selection biases arising from either cleaning or matching. This first impression was confirmed by a thorough examination of the distribution of key variables in the initial and final (matched) samples.
6.3.2. Choice of explanatory variables
We now detail the choice of explanatory variables used in the empirical specification of our econometric model. This choice of variable has been made in accordance with the original theoretical framework proposed in Crépon et al. (1998), taking into account some of the changes introduced to this initial CDM framework by Griffith et al. (2006), and trying to circumvent some limitations of the latter.
The x _{1} vector of regressors used in the first—‘selection into R&D’—equation of System (1) includes proxy variables for a firm’s ability to rip profit from innovation (appropriability conditions), and for market conditions and other Schumpeterian determinants of innovation. Appropriability conditions are represented by two dummy variables which describe how firms protected (p.138) their inventions during the observation period. The first (formal protection) is equal to 1 if a firm used patents, design patterns, trademarks, or copyrights, and to 0 otherwise. The second variable (strategic protection) is equal to 1 if a firm relied on complexity of design, secrecy, or leadtime advantage on competitors, and to 0 otherwise. Market conditions are partly captured by a binary indicator of international competition (equal to 1 if a firm’s most significant market is international, and to 0 otherwise). This variable may also be a proxy for the degree of openness of a firm to the international market. Therefore, market conditions in a more Schumpeterian sense are more aptly captured, in each application of the model, by the sets of industry dummy variables presented in Table 6.1 (and based on the twodigit NACE codes). These industry dummies give an indication of a firm’s main market during the observation period, and may thus control for market characteristics such as concentration.
The other usual Schumpeterian determinant of innovation included in the x _{1} vector is firm size. It is measured by the number of employees two years before the year of the survey, and represented by a fivecategory variable: (1) less than 50 employees, (2) 50 to 99 employees, (3) 100 to 249 employees, (4) 250 to 999 employees, and (5) 1000 or more employees. In order to make identification of the model easier, we chose to exclude firm size from the second equation of System (1), i.e. the ‘R&D intensity equation’. The choice of size as an exclusion variable was dictated by the results of previous studies. For instance, Griffith et al. (2006) have shown that, in several European countries (and particularly in France), firm size influences the probability to do R&D, but not the amount of investment in R&D.
Finally, in accordance with the original framework of Crépon et al. (1998), the x _{1} vector also includes variables indicating to which extent innovation was demand pulled or technology pushed in the threedigit industry where a firm operated during the observation period. These variables are built using a question that is specific to French CIS surveys. In each case, three variables give the share of firms where innovation was weakly/mildly/strongly influenced by market (or technological) conditions, while a fourth variable indicating no influence at all is taken as the reference. By measuring these variables at the threedigit industry level rather than at the firm level, we get around a difficulty raised by most innovation surveys: the fact that only innovating firms answer questions pertaining to the nature of innovation.
The x _{2} vector used in the second—‘R&D intensity’—equation of System (1) includes the same variables as the x _{1} vector, except firm size. However, the x _{2} vector includes three variables which are only observed when firms report R&D expenditures, and which may be useful to further characterize the R&D process. These three additional variables are: (1) cooperation in innovation activities, (2) public funding, and (3) sources of information used in the innovation process. Cooperation is a dummy variable equal to 1 if a firm had (p.139) some cooperative agreements on innovation activities during the observation period, and equal to 0 otherwise. Public funding is captured by a set of three nonmutually exclusive binary variables: (1) local funding, which indicates funding from local or regional authorities, (2) national funding, which indicates funding from the national government, and (3) EU funding , which indicates funding from the EU, including that received through participation in a Framework Programme. Similarly, sources of information are represented by a set of dummy variables which includes internal sources (within the firm), internal sources within the group (if the firm belongs to a group), suppliers, customers, competitors, and public research (i.e. universities or government labs).
The x _{3} vector of regressors used in the third equation of System (1)—the ‘product innovation equation’—includes an endogenous variable and three exogenous variables. The endogenous variable is the log of R&D intensity (predicted from the previous stage of the model). The exogenous variables are appropriability conditions (captured by the same indicators of invention protection as above),^{5} firm size (defined in exactly the same way as above), and the set of industryspecific dummies already described above and presented in Table 6.1. The x _{4} vector of regressors used in the fourth equation of System (1)—the ‘process innovation equation’—includes exactly the same variables as x _{3}, plus an additional variable. This additional variable is the amount of investment in physical capital, measured by the log of investment intensity in t2 (t being the year the survey was conducted).^{6} The rationale for including this variable is that, by definition, process innovation involves changes in the production line, which may require the acquisition of new machinery and equipment. The variable is lagged by two years in order to reduce potential simultaneity problems, which may arise since innovation is reported over the last three years (i.e. innovation can occur in t2, t1, or t).
The x _{5} vector of regressors used in the fifth and final equation of System (1)—the ‘labour productivity equation’—includes two endogenous variables and three exogenous variables. The two endogenous variables are the indicators of product and process innovation used as dependent variables in the innovation equations. In the threestep estimation procedure, these variables are the predicted marginal probabilities from the previously estimated bivariate probit. In the twostep sequential procedure, we use the actual (observed) binary indicators. The exogenous variables are firm size (defined in the same way as above), our set of industryspecific dummy variables and a proxy for (p.140) physical capital. In order to leave less scope for potential endogeneity problems, this proxy for physical capital is different from the one included in x _{4}. In the labour productivity equation, we measure physical capital using (the log of) the book value of fixed assets per employee at t1 (t being the year the survey was conducted). The variable is lagged by one year in order to account for the stock of physical capital at the end of the (previous) year.
6.3.3. Comparison of the samples
We have now presented our data sources and explained which variables will be used in the econometric model. Before moving to the results of the analysis, we briefly compare the three matched samples. Table 6.2 gives summary statistics for all of our dependent and explanatory variables in all three matched samples. Although differences in mean are often significant, the CIS3 and CIS4 manufacturing samples look very similar on average. In particular, the value of our measure of labour productivity (computed using value added) is quite similar, even without the use of a deflator, and the difference is not statistically significant. The main differences between the two manufacturing samples concern innovation: there are more process innovators in CIS4 (45 per cent versus only 32 per cent in CIS3), and innovation seems to have become more demandpulled. The importance of invention protection (and more specifically strategic protection) has also grown from CIS3 to CIS4. Finally, international competition has intensified in CIS4.
Focusing only on CIS4 emphasizes the differences between manufacturing and services in the recent period. Differences in means are more systematically significant across manufacturing and services than they were across the two manufacturing samples. Computed as the ratio of turnover to employees, labour productivity appears to be much lower in the services (130 euros per worker versus roughly 200 euros in the CIS3 and CIS4 manufacturing samples). When it is computed using value added instead of turnover, labour productivity is actually a little higher in the services (62 euros per worker versus roughly 55 euros in the manufacturing samples), and the difference is statistically significant. The reason is naturally that turnoverbased labour productivity measures take into account the costs of materials which correspond to much higher shares in manufacturing than in the services. The valueaddedbased measures abstracting from such differences may be preferred and viewed as more consistent with the fact that France (like most Western European countries) is gradually becoming a servicesdriven economy.
A major difference between manufacturing and services is that the proportion of firms doing R&D continuously is much lower in the latter (14 per cent versus 33 per cent in the manufacturing industry). The proportion of firms receiving public support to innovate is also much lower in the services: for (p.141)
Table 6.2. Summary statistics for the final samples
CIS3 
CIS4 


Manufacturing 
Manufacturing 
Services 

Knowledge/Innovation: 

Continuous R&D engagement 
0.38*** 
0.33 
0.14*** 
R&D per employee 
7.41 
7.66 
7.50 
(for firms w/continuous R&D) 

Innovation (product and/or process) 
0.52*** 
0.57 
0.36*** 
Process innovation 
0.32*** 
0.45 
0.31*** 
Product innovation 
0.47*** 
0.43 
0.23*** 
Share of sales with new products 
0.24* 
0.23 
0.19*** 
(for firms with product innovation) 

Labour productivity (Turnover) 
182.64** 
209.75 
130.05*** 
Labour productivity (Value added) 
54.01 
55.05 
62.4*** 
Public Support: 

Local funding (1 if yes, 0 if no) 
0.06** 
0.05 
0.01*** 
National funding (1 if yes, 0 if no) 
0.16*** 
0.12 
0.04*** 
EU funding (1 if yes, 0 if no) 
0.05** 
0.04 
0.02*** 
Innovation was: 

Not demand pulled 
0.03*** 
0.05 
0.11*** 
Weakly demand pulled 
0.07*** 
0.04 
0.06*** 
Mildly demand pulled 
0.28*** 
0.21 
0.22*** 
Strongly demand pulled 
0.61*** 
0.70 
0.62*** 
Not technology pushed 
0.12*** 
0.15 
0.23*** 
Weakly technology pushed 
0.23*** 
0.19 
0.15*** 
Mildly technology pushed 
0.43*** 
0.37 
0.32*** 
Strongly technology pushed 
0.22*** 
0.29 
0.30*** 
Sources of information 

(for firms doing R&D or innovating): 

Internal sources (w/in the firm) 
0.47** 
0.50 
0.30*** 
Internal sources (w/in the group) 
0.22*** 
0.28 
0.17*** 
Suppliers 
0.28*** 
0.32 
0.19*** 
Customers 
0.42*** 
0.37 
0.20*** 
Competitors 
0.33*** 
0.23 
0.13*** 
Universities/Government labs 
0.11 
0.12 
0.05*** 
Appropriability conditions: 

Formal protection dummy variable 
0.45*** 
0.51 
0.29*** 
Strategic protection dummy variable 
0.28*** 
0.39 
0.17*** 
Cooperation dummy variable 
0.27** 
0.29 
0.16*** 
Other: 

International competition 
0.41*** 
0.54 
0.19*** 
Size: 〈50 employees 
0.29*** 
0.34 
0.39*** 
Size: 50–99 employees 
0.18 
0.20 
0.20 
Size: 100–250 employees 
0.20*** 
0.18 
0.16 
Size: 250–999 employees 
0.24 
0.23 
0.20*** 
Size: ≥1,000 employees 
0.08*** 
0.06 
0.05 
Observations 
3518 
4831 
3467 
Notes: The table displays average values. All monetary values are in thousands of euros.
CIS3 (CIS4) variables are observed over 1998–2000 (2002–2004), except R&D per employee, labour productivity and investment per employee, which are observed in 2000 (2002), and firm size (number of employees), which is measured in 1998 (2002).
***, **, and * indicate differences in means (with respect to CIS4 Manufacturing) significant at the 1%, 5%, and 10% levels, respectively.
To put it in a nutshell, although there are contrasts between the two French manufacturing samples observed at different periods, these contrasts are less important than the differences between the services and manufacturing industries in the same period. These differences should be kept in mind when studying innovation, as they highlight the specificity of the services.
6.4. Results
In this section, the results obtained with the threestep sequential procedure model are presented as a benchmark. Table 6.3 and Table 6.4 give these results for the French manufacturing industry observed in CIS3 and CIS 4 respectively, while Table 6.5 presents them for the French services industry observed in CIS4. Table 6.6 presents the correlation coefficients of the generalized residuals of the model for each one of these three samples. In the following two subsections, we comment on these results and compare them to those obtained with the potentially more efficient twostep estimation procedure (the full tables of results for this procedure are presented in the appendix).
6.4.1. Manufacturing in CIS3 and in CIS4
The first comparison that can be conducted on the basis of our estimations concerns the manufacturing industry, which is observed for both CIS3 (1998–2000) and CIS4 (2002–2004). We achieve this comparison by comparing the results presented in Table 6.3 with those presented in Table 6.4. We also compare both series of results with those obtained with the twostep estimation procedure and presented in Tables A6.1 and A6.2 in the appendix, respectively.
We first discuss the estimates of the selection and R&Dintensity equations. In the selection equation, firm size appears as a major determinant of the propensity to do R&D, which is consistent with both the Schumpeterian tradition and the empirical literature. Calculation of the marginal effects (p.143)
Table 6.3. Estimates of the threestep sequential model (CIS3, manufacturing)
R&D equations 
Innovation production function 
Productivity 


Selection 
Intensity 
Product 
Process 

Log(R&D intensity) 
— 
— 
1.18*** 
0.73*** 
— 
(0.08) 
(0.07) 

Product innovation 
— 
— 
— 
— 
0.05 
(0.09) 

Process innovation 
— 
— 
— 
— 
0.41*** 
(0.12) 

Log(Physical capital intensity) 
— 
— 
— 
0.11*** 
0.13*** 
(0.02) 
(0.01) 

International competition 
0.40*** 
0.44*** 
— 
— 
— 
(0.06) 
(0.08) 

Cooperation 
— 
0.25*** 
— 
— 
— 
(0.07) 

Appropriability conditions 

Formal protection 
0.96*** 
0.30*** 
0.53*** 
0.16** 
— 
(0.06) 
(0.09) 
(0.06) 
(0.06) 

Strategic protection 
0.73*** 
0.30*** 
0.39*** 
0.32*** 
— 
(0.06) 
(0.08) 
(0.08) 
(0.07) 

Funding 

Local funding 
— 
0.03 
— 
— 
— 
(0.12) 

National funding 
— 
−0.14 
— 
— 
— 
(0.08) 

EU funding 
— 
0.46*** 
— 
— 
— 
(0.13) 

Firm size (ref.: 〈 50 employees) 

50 to 99 employees 
0.28*** 
— 
0.22*** 
0.13 
−0.09*** 
(0.08) 
(0.08) 
(0.08) 
(0.02) 

100 to 249 employees 
0.39*** 
— 
0.09 
0.23*** 
−0.15*** 
(0.08) 
(0.08) 
(0.08) 
(0.03) 

250 to 999 employees 
0.78*** 
— 
0.36*** 
0.27*** 
−0.13*** 
(0.08) 
(0.09) 
(0.08) 
(0.02) 

≥ 1,000 employees 
0.92*** 
— 
0.25** 
0.38*** 
−0.04 
(0.11) 
(0.12) 
(0.12) 
(0.05) 

Twodigit industry 
0.000 
0.000 
0.000 
0.000 
0.000 
Demand pulled/Techn. pushed 
0.833 
0.000 
— 
— 
— 
Sources of information 
— 
0.072 
— 
— 
— 
Sigma 
1.29*** 
— 
— 
— 

(0.04) 
Rho 
0.35*** 
0.47*** 
— 
(0.05) 
(0.03) 

GoodnessofFit 
Loglikelihood: 3667.18 
Loglikelihood: 2760.66 
Adj. R² = 0.27 
Test of ‘β = 0’: 523.60*** 
Test of ‘β = 0’: 1644.92*** 
Notes: * Significant at the 10% level,
** Significant at the 5% level,
*** Significant at the 1% level
Robust standard errors in parentheses.
For the sake of concision, we report only the pvalues of a test of the joint significance of the twodigit industry dummy variables, ‘demand pulled/technology pushed’ indicators, and ‘sources of information’ indicators.
Physical capital is measured by the log of investment intensity at time t2 in the process equation, and by the log of book value at t1 in the productivity equation.
Table 6.4. Estimates of the threestep sequential model (CIS4, manufacturing)
R&D equations 
Innovation production function 
Productivity 


Selection 
Intensity 
Product 
Process 

Log(R&D intensity) 
— 
— 
1.27*** 
0.90*** 
— 
(0.07) 
(0.06) 
— 

Product innovation 
— 
— 
— 
— 
−0.08 
(0.13) 

Process innovation 
— 
— 
— 
— 
0.45*** 
(0.16) 

Log(Physical capital intensity) 
— 
— 
— 
0.08*** 
0.10*** 
(0.02) 
(0.01) 

International competition 
0.51*** 
0.51*** 
— 
— 
— 
(0.05) 
(0.10) 

Cooperation 
— 
0.36*** 
— 
— 
— 
(0.08) 

Appropriability conditions 

Formal protection 
0.59*** 
0.43*** 
0.09 
−0.16*** 
— 
(0.05) 
(0.10) 
(0.05) 
(0.06) 

Strategic protection 
0.67*** 
0.61*** 
−0.17** 
−0.01 
— 
(0.05) 
(0.09) 
(0.06) 
(0.06) 

Funding 

Local funding 
— 
−0.03 
— 
— 
— 
(0.14) 

National funding 
— 
0.07 
— 
— 
— 
(0.09) 

EU funding 
— 
0.40*** 
— 
— 
— 
(0.14) 

Firm size (ref.: 〈 50 employees) 

50 to 99 employees 
0.18*** 
— 
0.05 
0.01 
−0.05** 
(0.07) 
(0.06) 
(0.06) 
(0.02) 

100 to 249 employees 
0.39*** 
— 
0.07 
0.05 
−0.10*** 
(0.07) 
(0.07) 
(0.06) 
(0.02) 

250 to 999 employees 
0.63*** 
— 
0.21*** 
0.05 
−0.06** 
(0.06) 
(0.07) 
(0.06) 
(0.02) 

≥ 1,000 employees 
1.07*** 
— 
0.35*** 
0.34*** 
0.02 
(0.10) 
(0.12) 
(0.11) 
(0.04) 

Twodigit industry 
0.000 
0.000 
0.000 
0.000 
0.000 
Demand pulled/Techn. pushed 
0.014 
0.000 
— 
— 
— 
Sources of information 
— 
0.053 
— 
— 
— 
Sigma 
— 
1.46*** 
— 
— 
— 
(0.04) 
Rho 
0.28*** 
0.44*** 
— 
(0.06) 
(0.03) 

Goodnessoffit 
Loglikelihood: 5061.75 
Loglikelihood: 4318.03 
Adj. R² = 0.21 
Test of ‘β = 0’: 444.41*** 
Test of ‘β = 0’: 2205.76*** 
Notes: * Significant at the 10% level,
** Significant at the 5% level,
*** Significant at the 1% level.
Robust standard errors in parentheses.
For the sake of concision, we report only the pvalues of a test of the joint significance of the twodigit industry dummy variables, ‘demand pulled/technology pushed’ indicators, and ‘sources of information’ indicators.
Physical capital is measured by the log of investment intensity at time t2 in the process equation, and by the log of book value at t1 in the productivity equation.
Table 6.5. Estimates of the threestep sequential model (CIS4, services)
R&D equations 
Innovation production function 
Productivity 


Selection 
Intensity 
Product 
Process 

Log (R&D intensity) 
— 
— 
0.32*** 
0.20*** 
— 
(0.04) 
(0.03) 
— 

Product innovation 
— 
— 
— 
— 
0.27 
(0.45) 

Process innovation 
— 
— 
— 
— 
0.27 
(0.52) 

Log (Physical capital intensity) 
— 
— 
— 
−0.01 
0.18*** 
(0.02) 
(0.01) 

International competition 
0.44*** 
0.90*** 
— 
— 
— 
(0.07) 
(0.20) 

Cooperation 
— 
0.07 
— 
— 
— 
(0.17) 

Appropriability conditions 

Formal protection 
0.71*** 
0.94*** 
0.28*** 
0.26*** 
— 
(0.07) 
(0.23) 
(0.08) 
(0.07) 

Strategic protection 
0.89*** 
1.11*** 
0.49*** 
0.50*** 
— 
(0.07) 
(0.22) 
(0.09) 
(0.09) 

Funding 

Local funding 
— 
−0.40 
— 
— 
— 
(0.34) 

National funding 
— 
0.77*** 
— 
— 
— 
(0.21) 

EU funding 
— 
−0.27 
— 
— 
— 
(0.31) 

Firm size (ref.: 〈 50 employees) 

50 to 99 employees 
0.23*** 
— 
0.16* 
0.17** 
−0.07* 
(0.08) 
(0.09) 
(0.08) 
(0.04) 

100 to 249 employees 
0.25*** 
— 
0.17* 
0.23*** 
−0.08** 
(0.09) 
(0.09) 
(0.08) 
(0.04) 

250 to 999 employees 
0.39*** 
— 
0.35*** 
0.40*** 
−0.15*** 
(0.09) 
(0.09) 
(0.08) 
(0.04) 

≥ 1,000 employees 
0.83*** 
— 
0.77*** 
0.69*** 
−0.29*** 
(0.14) 
(0.13) 
(0.12) 
(0.05) 

Twodigit industry 
0.069 
0.050 
0.148 
0.141 
0.000 
Demand pulled/Techn. pushed 
0.005 
0.000 
— 
— 
— 
Sources of information 
— 
0.503 
— 
— 
— 
Sigma 
— 
2.05** (0.14) 
— 
— 
— 
Rho 
0.62*** 
0.65*** 
— 
(0.07) 
(0.03) 

Goodnessoffit 
Loglikelihood: 1987.65 
Loglikelihood: 2177.52 
Adj. R² = 0.42 
Test of ‘β = 0’: 233.42*** 
Test of ‘β = 0’: 747.34*** 
Notes: * Significant at the 10% level,
** Significant at the 5% level,
*** Significant at the 1% level.
Robust standard errors in parentheses.
For the sake of concision, we report only the pvalues of a test of the joint significance of the 2digit industry dummy variables, ‘demand pulled/technology pushed’ indicators, and ‘sources of information’ indicators.
Physical capital is measured by the log of investment intensity at time t2 in the process equation, and by the log of book value at t1 in the productivity equation.
Table 6.6. Correlation coefficients of generalized residuals for each sample
CIS3 
CIS4, Manufacturing 
CIS4, Services 


1 
1 
1 

0.18*** 
1 
0.23*** 
1 
0.23*** 
1 

0.13*** 
0.34*** 
1 
0.34*** 
0.11*** 
1 
0.22*** 
0.33*** 
1 

0.06** 
0.19*** 
0.26*** 
1 
0.20*** 
0.02 
0.25*** 
1 
0.14*** 
0.34*** 
0.39*** 
1 

0.19*** 
−0.01 
−0.02 
−0.03* 
1 
0.01 
0.17*** 
0.01 
−0.05*** 
1 
0.17*** 
0.01 
−0.03 
−0.02 
1 
Notes: (1) R&D propensity,
(2) R&D intensity,
(3) Product innovation,
(4) Process innovation,
(5) Labour productivity.
* Significant at the 10% level;
** Significant at the 5% level;
*** Significant at the 1% level.
Moreover, manufacturing firms that are better able to protect their inventions or innovations are also more likely (1) to do R&D continuously and (2) to invest more in R&D. In CIS3 and CIS4, both formal and strategic means of protection are associated with a higher propensity to do R&D and with a higher R&D intensity. These results are consistent with the theoretical literature, which suggests that firms will not start innovating (and will not invest in R&D) if they cannot protect the output of their innovation in a way that guarantees a higher profit. Finally, international competition and cooperation are positively associated with a higher R&D intensity in both samples, but the magnitudes of the effects are stronger in CIS4 than in CIS3. International competition can be seen here as a proxy for openness to the international market rather than a proxy for competition in the strict sense. Our results suggest that French firms that have a higher international exposure may also invest more in R&D (and our summary statistics suggest that this exposure has increased between CIS3 and CIS4).
We now discuss the estimates of the ‘knowledge production function’ equations. The most important result concerns the effect of the endogenous explanatory variable ‘log of R&D intensity’. This variable is significant in the product and process innovation equations, in both CIS3 and CIS4 manufacturing samples. Calculation of the marginal effects shows that, in CIS3, a oneunit increase in the logR&D intensity results in a 47 per cent (25 per cent) increase in the probability of doing product (process) innovation, with a standard deviation of 3 per cent (2 per cent). Similarly, in CIS4, a oneunit increase in the logR&D intensity results in a 50 per cent (36 per cent) increase in the probability of doing product (process) innovation, with a standard deviation of 2 per cent (2 per cent). Thus, in both manufacturing samples, the magnitude of the effect of R&D is stronger on product than on process innovation.^{7}
The twostep estimation procedure confirms the above results on the knowledge production function. In both manufacturing samples (see Tables A6.1 (p.148) and A6.2 in the appendix), we find a significantly positive effect of R&D on product and process innovation. The magnitude of this effect is always stronger on product than on process innovation.^{8} With the twostep estimation procedure, the marginal effect of the logR&D intensity on product (process) innovation is equal to 49 per cent (25 per cent) in the CIS3 sample and to 50 per cent (34 per cent) in the CIS4 manufacturing sample. These values are very close to those obtained with our benchmark estimates.^{9}
Finally, we turn to our most important result: the effect of each type of innovation on labour productivity. According to our benchmark threestep sequential estimates, process innovation appears as the main driver of labour productivity in both manufacturing samples, whereas the impact of product innovation is never significant. In the CIS3 (CIS4) manufacturing sample, being a process innovator results in a 41 per cent (45 per cent) increase in the log of labour productivity, with a standard deviation of 12 per cent (15 per cent). The twostep estimation procedure confirms these results, showing that product innovation has no significant impact on labour productivity, and that only process innovation matters. The estimates obtained with this procedure are however slightly weaker than those obtained with the threestep procedure: in the CIS3 (CIS4) manufacturing sample, being a process innovator results in a 35 per cent (38 per cent) increase in the log of labour productivity, with a standard deviation of 10 per cent (10 per cent).
The dominance of process innovation as a driver of productivity in both manufacturing samples is in line with the findings of previous studies (Griffith et al., 2006; Raffo et al., 2008). However, given the difficulties mentioned in Section 6.2.1 (especially the potentially strong correlation between product and process innovation), this result deserves to be examined more closely. A look at Table 6.6 suggests that such a correlation might be at work here, because unobserved factors may affect both types of innovation in the same direction. Indeed, the correlation coefficient between the generalized residuals of the product and process equations is significantly positive in both manufacturing samples. This can also be seen in the fact that the correlation coefficient of the errors of the bivariate probit, estimated as part of the likelihood function in both procedures, is also significantly positive.^{10} We therefore conduct a more thorough examination of the respective impact of product and process innovation in Section 6.4.3.
(p.149) 6.4.2. Manufacturing and services in CIS4
We now turn to the comparison between the CIS4 manufacturing and services samples. To achieve this comparison, we compare the results presented in Table 6.5 with those already presented in Table 6.4. As in Section 6.4.1, we also compare both series of results with those obtained with the twostep estimation procedure, which are presented in Tables A6.2 and A6.3 in the appendix.
We first discuss the estimates of the selection and R&Dintensity equations. This discussion will be brief, as these results are essentially the same as those observed when comparing both manufacturing samples across time. First of all, we find that size is as important a determinant of the propensity to do R&D in the services as it was in the manufacturing industry. However, while the probability of doing R&D continuously still increases with firm size, the marginal effects are much smaller in the services industry. From the smallest to the largest size, marginal effects rise from 0.04 (with a standard deviation of 0.02) in the first two size classes, to 0.07 (0.02) and ultimately 0.20 (0.04) for the largest firms. Moreover, as in the manufacturing industry, service firms that are better able to protect their inventions or innovations are also more likely (1) to do R&D continuously and (2) to invest more in R&D. The estimated effects of both formal and strategic means of protection are significantly positive in both R&D equations. Finally, as in the manufacturing industry, international competition and cooperation are both positively associated with a higher R&D intensity.
An interesting contrast appears between manufacturing and service firms: whether in CIS3 or in CIS4, manufacturing firms are likely to invest more in R&D when they receive funding from the EU. Other types of public support do not seem to matter for these firms. By contrast, service firms are likely to invest more in R&D when they receive funding from their national government, whereas other types of public support do not seem to matter. It may be that, because the main markets of service firms are national rather than international, these firms seek support, when doing R&D, from national rather than supranational authorities that are less likely to be interested in their performances. It must also be noted from Table 6.2 that local and EUlevel funding are far from widespread in the services industry, which may be what drives this result.
We now discuss the estimates of the ‘knowledge production function’ equations. In the service industries as in the manufacturing industries, the effect of the logR&D intensity is significant in both innovation equations, but stronger in the product innovation equation than in the process innovation equation. In the former (latter), the marginal effect of the logR&D intensity is equal to 9 per cent (7 per cent), with a standard deviation of 1 per cent (p.150) (1 per cent). Both figures are considerably smaller than those computed in the CIS4 manufacturing sample. The estimates obtained with the twostep procedure (see Table A6.3 in the appendix) lead to the same marginal effects. These results incidentally raise the nagging question of exactly what is product/process innovation in the services industry. Our findings suggest that in both cases, R&D does matter for innovation in the services, but is certainly not the only key resource.
We finally turn to the discussion of the effect of each type of innovation in the labour productivity equation. Here, a striking contrast appears between the estimates obtained in the manufacturing industry (where productivity appeared to be driven by process innovation) and those obtained in the services industry. In the latter, neither product nor process innovation appears to have a significant impact on labour productivity. A possible explanation for the nonsignificance of process innovation could be found in the measurement issue highlighted in Section 6.2.1.
However, a comparison with the twostep estimation procedure points towards a more precise interpretation. According to the twostep procedure, productivity in the services would actually be driven by product (rather than process) innovation: as can be seen in Table A6.3 in the appendix, being a product innovator is associated with a significant 39 per cent increase in labour productivity. These diverging conclusions may come from the fact that the threestep estimation procedure relies on the predicted values of the innovation indicators, whereas the twostep procedure relies on the actual, observed indicators. What they suggest, however, is that, in the services industry, both types of innovation are likely to be concurrent, up to the point that they both measure ‘overall’ innovation, as explained in Section 6.2.1. Depending on the estimation procedure, the positive effect of product innovation dominates or absorbs the effect of process innovation.
This interpretation is reinforced by the fact that unobserved factors may affect product and process in the same direction, as can be seen in Table 6.6. The correlation coefficient between the generalized residuals of the product and process equations is significantly positive and much higher (0.39) in the services sample than in the manufacturing samples (0.26 in CIS3 and 0.25 in CIS4). Similarly, the correlation coefficient of the errors of the bivariate probit model is higher in the services sample (around 0.64 or 0.65, depending on which estimation procedure is used) than it was in the manufacturing samples. The abovementioned interpretation therefore deserves a closer examination, which is conducted in Section 6.4.3 through sensitivity analyses. (p.151)
6.4.3. Sensitivity analyses
The main result of our analysis can be summarized as follows: in the manufacturing industries and across both periods, labour productivity appears primarily driven by process innovation, while in the services industry, if anything, it would be by product innovation. In view of the difficulties in disentangling the productivity impact of both types of innovation indicators, it could in fact be that they both approximately proxy for ‘overall’ innovation. This would explain why productivity appears to be sometimes correlated with product innovation or with process innovation, and even sometimes uncorrelated with both. In this section, we examine the robustness of these results by estimating variants of our model where we either (1) neglect one type of innovation, or (2) simply rely on an indicator of overall innovation.
In a first variant, the knowledge production function is simply the product innovation equation, specified as a probit model. We neglect process innovation, as if all innovation were actually product innovation. (This model is therefore an alternative specification of the original CDM model by Crépon et al. 1998.) In a second variant of the model, the knowledge production function is simply the process innovation equation, again specified as a probit model. This time, product innovation is neglected, as if all innovation were actually process innovation. In a third and final variant of the model, the knowledge production is specified as a probit model where the dependent variable is an indicator which equals 1 if a firm innovates in product and/or process, and 0 otherwise. Each variant encompasses only one knowledge production equation, which means that System (1) collapses to four equations instead of five.
Table 6.7 presents the key estimates, obtained with the threestep procedure, for: (1) our original model (i.e. the full model with both process and product innovation), (2) the model with only product innovation, (3) the model with only process innovation, and (4) the model with a single indicator of overall innovation. The key estimates for each model are (1) the effect of logR&D intensity on innovation and (2) the effect of innovation on labour productivity. We discuss these key estimates and compare them to those obtained with the twostep estimation procedure (presented in Table A6.4 in the appendix).
Table 6.7 provides strong support for the assumption that our indicators of product and process innovation are actually both measuring ‘overall’ innovation. First of all, in the model that only considers product innovation, its predicted probability always has a positive effect on labour productivity. According to the threestep procedure estimates, being a product innovator is associated with an increase in labour productivity of about 30 per cent, 27 per cent, and 60 per cent in the CIS3 manufacturing sample, CIS4 manufacturing sample, and CIS4 services sample respectively. Estimating the same model with the twostep procedures yields estimates equal to 29 per cent, 26 (p.152)
Table 6.7. Key estimates of the CDM model with alternative versions of the knowledge production function
Knowledge production function = product innovation, process innovation (bivariate probit) 


CIS3 
CIS4, Manufacturing 
CIS4, Services 

Prod. 
Proc. 
LP 
Prod. 
Proc. 
LP 
Prod. 
Proc. 
LP 

Log (R&D 
1.18*** 
0.73*** 
1.27*** 
0.90*** 
0.32*** 
0.20*** 

intensity) 
(0.08) 
(0.07) 
(0.07) 
(0.06) 
(0.04) 
(0.03) 

Product 
— 
— 
0.05 
— 
— 
−0.08 
— 
— 
0.27 
innovation 
(0.09) 
(0.13) 
(0.45) 

Process 
— 
— 
0.41*** 
— 
— 
0.45*** 
— 
— 
0.27 
innovation 
(0.12) 
(0.16) 
(0.52) 
Knowledgeproduction function = product innovation (simple probit) 

CIS3 
CIS4, Manufacturing 
CIS4, Services 

Prod. 
LP 
Prod. 
LP 
Prod. 
LP 

Log (R&D 
1.27*** 
— 
1.31*** 
— 
0.32*** 
— 
intensity) 
(0.07) 
(0.06) 
(0.03) 

Product 
— 
0.30*** 
— 
0.27*** 
— 
0.60*** 
innovation 
(0.04) 
(0.03) 
(0.06) 
Knowledgeproduction function = process innovation (simple probit) 

CIS3 
CIS4, Manufacturing 
CIS4, Services 

Proc. 
LP 
Proc. 
LP 
Proc. 
LP 

Log (R&D 
0.71*** 
— 
0.90*** 
— 
0.18*** 
— 
intensity) 
(0.06) 
(0.06) 
(0.03) 

Process 
— 
0.49*** 
— 
0.35*** 
— 
0.59*** 
innovation 
(0.06) 
(0.04) 
(0.06) 
Knowledgeproduction function = ‘overall’ innovation (simple probit) 

CIS3 
CIS4, Manufacturing 
CIS4, Services 

Inno. 
LP 
Inno. 
LP 
Inno. 
LP 

Log (R&D 
1.45*** 
— 
1.48*** 
— 
0.24*** 
— 
intensity) 
(0.08) 
(0.08) 
(0.03) 

‘Overall’ 
— 
0.31*** 
— 
0.28*** 
— 
0.53*** 
innovation 
(0.04) 
(0.03) 
(0.06) 
Notes: * Significant at the 10% level,
** Significant at the 5% level,
*** Significant at the 1% level
Robust standard errors in parentheses.
Prod. = Product innovation, Proc. = Process innovation, Inno. = ‘overall’ innovation, LP = Labour productivity.
Similarly, in the model that only considers process innovation, its predicted probability has a positive effect on labour productivity in all three samples. According to the threestep procedure estimates, being a process innovator increases labour productivity by 49 per cent, 35 per cent, and 59 per cent in the CIS3 manufacturing, CIS4 manufacturing, and CIS4 services samples respectively. Again, the twostep procedures yield estimates that are roughly stable in the manufacturing samples (41 per cent and 37 per cent for CIS4 and CIS3 respectively), but definitely lower in the services sample (40 per cent).
Last but not least, the model where the knowledge production function models ‘overall’ innovation (without distinguishing between product and process) shows that the predicted innovation probability always has a positive effect on labour productivity. The estimated elasticities effects obtained with the benchmark threestep procedure are equal to 31 per cent, 28 per cent, and 53 per cent in the CIS3 manufacturing, CIS4 manufacturing, and CIS4 services samples respectively. The estimates obtained when estimating the model with the twostep procedure are very similar in the manufacturing samples (29 per cent and 25 per cent for CIS3 and CIS4 respectively) and somewhat lower in the services samples (40 per cent).
In a nutshell, our sensitivity analysis suggests that our indicators of product and process innovation tend to capture the overall effect of innovation. This seems to be especially the case in the services industry, where the high estimates effect points at a stronger divide between innovating firms and noninnovators. In the end, disentangling the respective effects of product and process innovation could well prove to be a harder task than most studies consider. Experiments with an interaction term in the threestep procedure showed that the problem was actually pervasive, and that it might be preferable to consider only overall innovation.^{11}
6.5. Conclusion
In this chapter, we examined the effect of innovation on labour productivity in France, using a CDMtype framework that accounts for research activities and for both product and process innovation. We estimated a multipleequation (p.154) econometric model using threestep and twostep estimation procedures based on maximum likelihood. The model was specified as a system of five equations. The first two accounted for the propensity to do R&D and for the amount of investment in R&D. The next two equations modelled the knowledge production function, distinguishing between product and process innovation. The final equation measured the impact of each type of innovation on labour productivity.
We estimated this model on the third and fourth waves of the French component of the Community Innovation Survey (CIS3 and CIS4). Our main results point to process innovation as the main driver of labour productivity in the manufacturing industry in both waves. However, a more careful examination, through sensitivity analyses, suggests that our indicators of product and process innovation both account for ‘overall’ innovation, especially in the services industry. When a given type of innovation is singled out, it has a positive effect on labour productivity in all periods and in all samples. Similarly, when an indicator of overall innovation is built, its effect on labour productivity is always positive, in all industries. Thus, disentangling the effects of product and process innovation may sometimes be more difficult than is usually thought. In such a situation, relying on a single general indicator of innovation might be preferable.
(p.155) Appendix
Table A6.1. Estimates of the twostep sequential model (CIS3, manufacturing)
R&D equations 
Innovation production function 
Productivity 


Selection 
Intensity 
Product 
Process 

Log(R&D intensity) 
— 
— 
1.24*** 
0.74*** 
— 
(0.07) 
(0.06) 

Product innovation 
— 
— 
— 
— 
0.08 
(0.08) 

Process innovation 
— 
— 
— 
— 
0.35*** 
(0.10) 

Log(Physical capital intensity) 
— 
— 
— 
0.14*** 
0.08*** 
(0.02) 
(0.01) 

International competition 
0.40*** 
0.44*** 
— 
— 
— 
(0.06) 
(0.08) 

Cooperation 
— 
0.25*** 
— 
— 
— 
(0.07) 

Appropriability conditions 

Formal protection 
0.96*** 
0.30*** 
0.54*** 
0.16** 
— 
(0.06) 
(0.09) 
(0.06) 
(0.06) 

Strategic protection 
0.73*** 
0.30*** 
0.32*** 
0.24*** 
— 
(0.06) 
(0.08) 
(0.07) 
(0.06) 

Funding 

Local funding 
— 
0.03 
— 
— 
— 
(0.12) 

National funding 
— 
−0.14 
— 
— 
— 
(0.08) 

EU funding 
— 
0.46*** 
— 
— 
— 
(0.13) 

Firm size (ref.: 〈 50 employees) 

50 to 99 employees 
0.28*** 
— 
0.23*** 
0.16** 
−0.10*** 
(0.08) 
(0.07) 
(0.08) 
(0.02) 

100 to 249 employees 
0.39*** 
— 
0.09 
0.21*** 
−0.12*** 
(0.08) 
(0.07) 
(0.08) 
(0.03) 

250 to 999 employees 
0.78*** 
— 
0.37*** 
0.23*** 
−0.07** 
(0.08) 
(0.07) 
(0.08) 
(0.03) 

≥ 1,000 employees 
0.92*** 
— 
0.26** 
0.33*** 
0.01 
(0.11) 
(0.12) 
(0.11) 
(0.04) 

Twodigit industry 
0.000 
0.000 
0.000 
0.000 
0.000 
Demand pulled/Techn. pushed 
0.833 
0.000 
— 
— 
— 
Sources of Information 
— 
0.072 
— 
— 
— 
Sigma 
— 
1.29*** 
— 
— 
0.49*** 
(0.04) 
(0.02) 
Rho 
0.35*** 
Rho (Product, Process) = 0.47*** (0.03) 
Rho (Product, Productivity) = −0.23*** (0.08) 

(0.05) 

Rho (Process, Productivity) = −0.49*** (0.10) 

Goodnessoffit 
Loglikelihood: −3667.18 
Loglikelihood: −4984.86 
Test of ‘β = 0’: 523.60*** 
Test of ‘β = 0’: 1241.70*** 
Notes: * Significant at the 10% level;
** Significant at the 5% level;
*** Significant at the 1% level.
Robust standard errors in parentheses.
For the sake of concision, we report only the pvalues of a test of the joint significance of the twodigit industry dummy variables, ‘demand pulled/technology pushed’ indicators, and ‘sources of information’ indicators.
Physical capital is measured by the log of investment intensity at time t2 in the process equation, and by the log of book value at t1 in the productivity equation.
Table A6.2. Estimates of the twostep sequential model (CIS4, manufacturing)
R&D equations 
Innovation production function 
Productivity 


Selection 
Intensity 
Product 
Process 

Log(R&D intensity) 
— 
— 
1.29*** 
0.85*** 
— 
(0.06) 
(0.06) 
— 

Product innovation 
— 
— 
— 
— 
−0.04 
(0.08) 

Process innovation 
— 
— 
— 
— 
0.38*** 
(0.10) 

Log(Physical capital intensity) 
— 
— 
— 
0.13*** 
0.07*** 
(0.02) 
(0.01) 

International competition 
0.51*** 
0.51*** 
— 
— 
— 
(0.05) 
(0.10) 

Cooperation 
— 
0.36*** 
— 
— 
— 
(0.08) 

Appropriability conditions 

Formal protection 
0.59*** 
0.43*** 
0.08 
−0.09* 
— 
(0.05) 
(0.10) 
(0.05) 
(0.06) 

Strategic protection 
0.67*** 
0.61*** 
−0.19*** 
−0.04 
— 
(0.05) 
(0.09) 
(0.06) 
(0.06) 

Funding 

Local funding 
— 
−0.03 
— 
— 
— 
(0.14) 

National funding 
— 
0.07 
— 
— 
— 
(0.09) 

EU funding 
— 
0.40*** 
— 
— 
— 
(0.14) 

Firm size (ref.: 〈 50 employees) 

50 to 99 employees 
0.18*** 
— 
0.06 
−0.0004 
−0.04* 
(0.07) 
(0.06) 
(0.06) 
(0.02) 

100 to 249 employees 
0.39*** 
— 
0.06 
0.02 
−0.08*** 
(0.07) 
(0.06) 
(0.06) 
(0.02) 

250 to 999 employees 
0.63*** 
— 
0.21*** 
0.01 
−0.02 
(0.06) 
(0.06) 
(0.06) 
(0.02) 

≥ 1,000 employees 
1.07*** 
— 
0.36*** 
0.28** 
0.07* 
(0.10) 
(0.12) 
(0.11) 
(0.04) 

Twodigit industry 
0.000 
0.000 
0.000 
0.000 
0.000 
Demand pulled/Techn. pushed 
0.014 
0.000 
— 
— 
— 
Sources of information 
— 
0.053 
— 
— 
— 
Sigma 
— 
1.46*** 
— 
— 
0.50*** 
(0.04) 
(0.02) 
Rho 
0.28*** 
Rho (Product, Process) = 0.43*** (0.03) 
Rho (Product, Productivity) = 0.08 (0.06) 

(0.06) 

Rho (Process, Productivity) = 0.52*** (0.10) 

Goodnessoffit 
Loglikelihood: 5061.75 
Loglikelihood: 7509.18 
Test of ‘β = 0’: 444.41*** 
Test of ‘β = 0’: 1662.29*** 
Notes: * Significant at the 10% level;
** Significant at the 5% level;
*** Significant at the 1% level.
Robust standard errors in parentheses.
For the sake of concision, we report only the pvalues of a test of the joint significance of the 2digit industry dummy variables, ‘demand pulled/technology pushed’ indicators, and ‘sources of information’ indicators.
Physical capital is measured by the log of investment intensity at time t2 in the process equation, and by the log of book value at t1 in the productivity equation.
Table A6.3. Estimates of the twostep sequential model (CIS4, services)
R&D equations 
Innovation production function 
Productivity 


Selection 
Intensity 
Product 
Process 

Log(R&D intensity) 
— 
— 
0.34*** 
0.22*** 
— 
(0.03) 
(0.03) 
— 

Product innovation 
— 
— 
— 
— 
0.39*** 
(0.08) 

Process innovation 
— 
— 
— 
— 
0.09 
(0.09) 

Log(Physical capital intensity) 
— 
— 
— 
−0.005 
0.15*** 
(0.02) 
(0.01) 

International competition 
0.44*** 
0.90*** 
— 
— 
— 
(0.07) 
(0.20) 

Cooperation 
— 
0.07 
— 
— 
— 
(0.17) 

Appropriability conditions 

Formal protection 
0.71*** 
0.94*** 
0.29*** 
0.26*** 
— 
(0.07) 
(0.23) 
(0.06) 
(0.07) 

Strategic protection 
0.89*** 
1.11*** 
0.43*** 
0.44*** 
— 
(0.07) 
(0.22) 
(0.07) 
(0.10) 

Funding 

Local funding 
— 
−0.40 
— 
— 
— 
(0.34) 

National funding 
— 
0.77*** 
— 
— 
— 
(0.21) 

EU funding 
— 
−0.27 
— 
— 
— 
(0.31) 

Firm size (ref.: 〈 50 employees) 

50 to 99 employees 
0.23*** 
— 
0.15** 
0.16** 
−0.07** 
(0.08) 
(0.06) 
(0.08) 
(0.03) 

100 to 249 employees 
0.25*** 
— 
0.19*** 
0.24*** 
−0.08*** 
(0.09) 
(0.07) 
(0.08) 
(0.03) 

250 to 999 employees 
0.39*** 
— 
0.38*** 
0.41*** 
−0.15*** 
(0.09) 
(0.06) 
(0.08) 
(0.03) 

≥ 1,000 employees 
0.83*** 
— 
0.74*** 
0.68*** 
−0.28*** 
(0.14) 
(0.10) 
(0.12) 
(0.04) 

Twodigit industry 
0.069 
0.050 
0.000 
0.007 
0.000 
Demand pulled/Techn. pushed 
0.005 
0.000 
— 
— 
— 
Sources of information 
— 
0.503 
— 
— 
— 
Sigma 
— 
2.05** 
— 
— 
0.57*** 
(0.14) 
(0.01) 
Rho 
0.62*** 
Rho (Product, Process) = 0.64*** (0.03) 
Rho (Product, Productivity) = −0.45*** (0.06) 

(0.07) 

Rho (Process, Productivity) = −0.27** (0.10) 

Goodnessoffit 
Loglikelihood: −1987.65 
Loglikelihood: −4775.80 
Test of ‘β = 0’: 233.42*** 
Test of ‘β = 0’: 1016.62*** 
Notes: * Significant at the 10% leve;
** Significant at the 5% level;
*** Significant at the 1% level.
Robust standard errors in parentheses. For the sake of concision, we report only the pvalues of a test of the joint significance of the 2digit industry dummy variables, ‘demand pulled/technology pushed’ indicators, and ‘sources of information’ indicators. Physical capital is measured by the log of investment intensity at time t2 in the process equation, and by the log of book value at t1 in the productivity equation.
Table A6.4. Key twostep estimates of the CDM model with alternative knowledge production functions
Knowledge production function = product innovation, process innovation (bivariate probit) 


CIS3 
CIS4, Manufacturing 
CIS4, Services 

Prod. 
Proc. 
LP 
Prod. 
Proc. 
LP 
Prod. 
Proc. 
LP 

Log(R&D 
1.24*** 
0.74*** 
1.29*** 
0.85*** 
0.34*** 
0.22*** 

intensity) 
(0.07) 
(0.06) 
(0.06) 
(0.06) 
(0.03) 
(0.03) 

Product 
— 
— 
0.08 
— 
— 
−0.04 
— 
— 
0.39*** 
innovation 
(0.08) 
(0.10) 
(0.08) 

Process 
— 
— 
0.35*** 
— 
— 
0.38*** 
— 
— 
0.09 
innovation 
(0.10) 
(0.10) 
(0.09) 
Knowledgeproduction function = product innovation (simple probit) 


CIS3 
CIS4, Manufacturing 
CIS4, Services 

Prod. 
LP 
Prod. 
LP 
Prod. 
LP 

Log(R&D 
1.29*** 
— 
1.29*** 
— 
0.34*** 
— 
intensity) 
(0.07) 
(0.06) 
(0.03) 

Product 
— 
0.29*** 
— 
0.26*** 
— 
0.48*** 
innovation 
(0.03) 
(0.03) 
(0.05) 
Knowledge production function = process innovation (simple probit) 


CIS3 
CIS4, Manufacturing 
CIS4, Services 

Proc. 
LP 
Proc. 
LP 
Proc. 
LP 

Log(R&D 
0.72*** 
— 
0.84*** 
— 
0.22*** 
— 
intensity) 
(0.06) 
(0.06) 
(0.03) 

Process 
— 
0.41*** 
— 
0.37*** 
— 
0.40*** 
innovation 
(0.04) 
(0.05) 
(0.06) 
Knowledgeproduction function = ‘overall’ innovation (simple probit) 


CIS3 
CIS4, Manufacturing 
CIS4, Services 

Inno. 
LP 
Inno. 
LP 
Inno. 
LP 

Log(R&D 
1.47*** 
— 
1.44*** 
— 
0.28*** 
— 
intensity) 
(0.08) 
(0.08) 
(0.03) 

‘Overall’ 
— 
0.29*** 
— 
0.25*** 
— 
0.40*** 
innovation 
(0.02) 
(0.02) 
(0.05) 
Notes: * Significant at the 10% level;
** Significant at the 5% level;
*** Significant at the 1% level.
Robust standard errors in parentheses.
Prod. = Product innovation, Proc. = Process innovation, Inno. = ‘overall’ innovation, LP = Labour productivity.
References
Bibliography references:
Chudnosky, D., Lopez, A. and Pupato, G. (2006), ‘Innovation and productivity in developing countries: A study of Argentine manufacturing firms’ behavior (1992–2001)’, Research Policy, 35: 266–288.
Crépon, B., Duguet, E. and Mairesse, J. (1998), ‘Research, innovation and productivity: An econometric analysis at the firm level’, Economics of Innovation and New Technology, 7: 115–158.
Griffith, R., Huergo, E., Mairesse, J. and Peters, B. (2006), ‘Innovation and productivity across four European countries’, Oxford Review of Economic Policy, 22(4): 483–498.
Griliches, Z. (1979), ‘Issues in assessing the contribution of research and development to productivity growth’, Bell Journal of Economics, 87(5): 537–64.
Hall, B.H., Lotti, F. and Mairesse, J. (2009), ‘Innovation and productivity in SMEs: Empirical evidence for Italy’, Small Business Economics, 33: 13–33.
Mairesse, J. and Jaumandreu, J. (2005), ‘Paneldata estimates of the production function and the revenue function: What difference does it make?’, The Scandinavian Journal of Economics, 107(4): 651–72.
Mairesse, J. and Mohnen, P. (2010), ‘Using innovation surveys for econometric analysis’, in B.H. Hall and N. Rosenberg (eds), Handbook of the Economics of Innovation, Vol. II. London: Burlington Academic Press, pp. 1129–1155.
Mairesse, J. and Mohnen, P. (2002), ‘Accounting for innovation and measuring innovativeness: An illustrative framework and an application’, American Economic Review, 92(2): 226–230.
Mairesse, J., Mohnen, P. and Kremp, E. (2005), ‘The importance of R&D and innovation for productivity: A reexamination in light of the 2000 French Innovation Survey’, Annales d'economie et de statistique (Contributions in memory of Zri Griliches, ed. Jacques Mairesse and Manuel Trajtenberg 79/80: 487–527.
Mohnen, P., Mairesse, J. and Dagenais, M. (2006), ‘Innovativity: A comparison across seven European countries’, Economics of Innovation and New Technology, 15(4/5): 391–413.
Raffo, J., Lhuillery, S. and Miotti, L. (2008), ‘Northern and southern innovativity: A comparison across European and Latin American countries’, European Journal of Development Research, 20(2): 219–239.
Roodman, D. (2009), ‘Estimating fully observed recursive mixedprocess models with cmp’, CDG Working Paper 168, Washington, DC: Center for Global Development.
Notes:
(^{1}) The system is recursive in the sense that the five equations are nested in three clearly defined stages, each stage being modelled as the determinant of the next one. Thus, the first two equations model firms’ R&D effort (first stage), which is a determinant of the ‘knowledge production function’ (second stage), represented by the next two equations. Innovation (the knowledge output of the second stage) is a determinant of firms’ labour productivity, modelled in the fifth equation using a standard production function framework (third stage).
(^{2}) Using a bivariate probit rather than two separate probits (as in Griffith et al., 2006; and Raffo et al., 2008) has two advantages. First, by allowing unobserved factors to jointly affect product and process innovation, the bivariate probit may do a better job of accounting for the potential correlation between both types of innovation. Second, the estimation of a bivariate probit may yield a gain in efficiency in the estimation of the CDM model.
(^{3}) For manufacturing, we had to leave aside the agrofood industry, which is surveyed in the CIS but not covered in the administrative data. For services, we had to leave aside three sectors that are surveyed in CIS4 but not covered in the administrative data: ‘Transport services’, ‘Trade’, and ‘R&D services’. The latter is so specific anyway that it deserves a separate study.
(^{4}) Note also that the distribution of manufacturing firms across industries does not change much from CIS3 to CIS4.
(^{5}) Our indicators of invention protection are used as instruments, in the sense that they are included in both x _{3} and x _{4}, but not in the labour productivity equation, where product and process innovation are endogenous regressors.
(^{6}) Investment intensity is defined here as the ratio of investment to the number of employees.
(^{7}) This difference could simply mean that product innovation requires more R&D than process innovation. However, process innovation generally implies the purchase of new machines and equipment. We controlled for this by including a proxy for physical capital in the ‘process innovation’ equation. This proxy is significant (with a positive effect) in both manufacturing samples, which can also explain why the effect of R&D appears to be weaker in the ‘process innovation’ equation.
(^{8}) Again, the proxy for physical capital included in the ‘process innovation’ equation is always significant, with a positive effect.
(^{9}) And so are their standard errors which, for the sake of clarity, we do not display here.
(^{10}) This correlation coefficient is equal to 0.47 in the CIS3 manufacturing sample according to both estimation procedures. It is equal to 0.43 or 0.44 in the CIS4 manufacturing sample, depending on which procedure is used.
(^{11}) We did not conduct similar experiments with the twostep procedure, because there are no clear rules on how to introduce (let alone instrument) an interaction term between two nonpredicted endogenous regressors in a FIML simultaneousequation model such as Model (2). We keep this issue for future research.