Innovation and productivity in knowledge intensive business services

Abstract

This paper studies empirically the relationship between the sources of knowledge, innovation and productivity in Knowledge Intensive Business Services using French micro data and sheds some new light on the production of knowledge and on effects of innovative output on firm productivity. Both an innovation function and a production function augmented with dummy endogenous innovation are estimated. Three estimators which control for endogeneity of the dummy innovation are employed: the first is a maximum likelihood estimator of the equations’ system while the other two are built in the instrumental variables framework. These estimators give useful complementary information because of the usual efficiency-robustness tradeoff comparing system-equations and single-equation estimators. We find that innovation is frequent in Knowledge Intensive Business Services and has a strong and positive effect on productivity. As in manufacturing, the main determinant of innovation is formal knowledge resulting from R&D or from acquisitions of equipment, patents or licenses.

Appendices

Appendix: the data

1.1 Industry coverage of French CIS

The first Community Innovation Survey, CIS1, was carried out in 1991 on the basis of first version of the Oslo Manual and covers only the manufacturing sector. Services have been included since the second wave of the survey, CIS2 which covers several non-personal market services (NACE 60, 61, 62, 64.2) and business services (NACE 72, 74.2), while CIS3 covers only computer and related activities (NACE 72), other business activities (NACE 74.2, 74.3) and post and telecommunications (NACE 64). For the estimation of the model using CIS3 data, the empirical delimitation of KIBS is thus: computer and related activities (NACE 72), Architectural and engineering activities (NACE 74.2), checking and analyses techniques (NACE 74.3) and telecommunications (NACE 64.2). The inclusion of telecommunications—which is not strictly speaking a business activity—is related to the technological and innovative contents of such activities (see Table 7).

Table 7 Sectors covered by the CIS surveys

1.2 Variable definitions

1.2.1 Innovation

Five estimations of the innovation function are performed corresponding to the different definitions of innovation:

• I₁: Binary variable (yes/no) for any kind of innovation (“either” innovation);

• I₂: Binary variable (yes/no) for product innovation;

• I₃: Binary variable (yes/no) for process innovation;

• I₄: Binary variable (yes/no) for technological innovation;

• I₅: Binary variable (yes/no) for non technological innovation.

The French CIS3 questionnaire does not contain any direct question relating to the introduction of a technological or a non technological innovation in the same way as it does for product and process innovation. However, when a firm introduces a product or a process innovation it has to declare whether this innovation is technologically new. We use this information to build technological and non technological innovation indicators.

1.2.2 Knowledge inputs

The CIS3 survey provides several variables that we label knowledge inputs:

$$ {\bf z}_{1}=( r^{1}, r^{2}, r^{3}, k^{1}, k^{2}, k^{3}, k^{4}, k^{5}, c, p) $$

with:

• r¹: R&D internal activity (yes/no) “All creative work undertaken within your enterprise on a systematic basis in order to increase the stock of knowledge, and the use of this stock of knowledge to devise new applications, such as new and improved products (goods/ services) and processes (including software research”;

• r²: R&D external subcontracted activity (yes/no) “Same activities as above, but performed by other companies (including other enterprises within the group) or other public orprivate research organisations”;

• r³: Other expenditures associated with knowledge acquisition (patents, licenses, training, etc.) (yes/no) “Advanced machinery, computer hardware specifically purchased to implement new or significantly improved products (goods/services) and/or processes. Purchase of rights to use patents and non-patented inventions, licenses, know-how, trademarks, software and other types of knowledge from others for use in your enterprise’s innovations. Internal or external training for your personnel directly aimed at the development and/or introduction of innovations”;

• k¹: Internal sources of knowledge (at firm level) (yes/no) “Within the enterprise”;

• k²: Internal sources of knowledge (at corporate level) (yes/no) “Other enterprises within the enterprise group”;

• k³: External sources of knowledge, linked to the market (suppliers, customers, competitors) (yes/no) “Suppliers of equipment, materials, components or software, Clients or customers, Competitors and other enterprises from the same industry”;

• k⁴: External sources of knowledge, issued from Public research (yes/no) “Universities or other higher education institutes, Government or private non-profit research institutes”;

• k⁵: Other External sources of knowledge (Conferences, Articles, Meetings) (yes/no) “Professional conferences, meetings, journals, Fairs, exhibitions”;

• c: Cooperation for innovation (yes/no) “Innovation co-operation means active participation in joint R&D and other innovation projects with other organisations (either other enterprises or noncommercial institutions). It does not necessarily imply that both partners derive immediate commercial benefit from the venture. Pure contracting out of work, where there is no active collaboration, is not regarded as co-operation. Did your enterprise have any co-operation arrangements on innovation activities with other enterprises or institutions during 1998–2000?”;

• p: Public fundings for innovation (yes/no) “Public funding includes financial support in terms of grants and loans, including a subsidy element, and loan guarantees. Ordinary payments for orders of public customers shall not be included. Did your enterprise receive any public financial support for innovation activities during the period 1998–2000?”.

1.2.3 Absorptive capacity factors

Absorptive capacity factors:

$$ {\bf z}_{2}=( l_{96},\tau ^{1},\tau ^{2},\tau ^{3},m,o,s) $$

with:

• l₉₆: the size of the firm measured through the number of employees in 1996;

• τ¹: the demand pull effect (yes/no) “For your enterprise, the introduction of an innovation is determined by the market pull”;

• τ²: the technology push effect (yes/no) “For your enterprise, the introduction of an innovation is determined by advances in technological knowledge”;

• τ³: the non-technological knowledge push effect “For your enterprise, the introduction of an innovation is determined by advances in non technological knowledge”;

• m : the market level of the firm (region, country, Europe, international) “What is your enterprise’s most significant market?”;

• o change of the status or organization of the firm (yes/no) “Creation, mergers, functional out-sourcing or new management methods implementation” ;

• s: the industry affiliation of the firm, according to the three digit NACE classification.

Using both kinds of inputs, our vector z = (z ₁ , z ₂ ) of inputs is defined as:

$$ {\bf z}={\bf (z}_{1}{\bf ,z}_{2}{\bf )}\,=\,( r^{1},r^{2},r^{3},k^{1},k^{2},k^{3},k^{4},,k^{5},c,p,l_{96},\tau ^{1},\tau ^{2},\tau ^{3},m,o,s) $$

1.2.4 Production function

• y: value added;

• l: number of employees;

• c: investments in physical capital.

Missing values due to the filter: some robustness checks

The CIS3 French questionnaire is structured in such a way that the firms that have not engaged into an innovative project (having innovated, having introduced organisational change or having ongoing or abandoned innovation activities) over the period have not to answer to a number of questions, typically relating to the inputs of innovation as R&D, machinery, internal and external sources of knowledge, etc. In our sample about 1/3 of all firms have missing values due to the filter for many relevant variables.

We deal with this issue by imputing the missing values. This is done by using probit regression with sample selection to estimate the likelihood of using R&D or another source of knowledge. As explanatory variables we use firm-specific variables available for all firms, such as size, sectoral dummies, affiliation with a corporation, export rate and market level. The imputation of non responses is performed by using the predicted probabilities obtained from such regressions using 0.5 as a discriminatory threshold (see also Sect. 5.3).

In order to provide some robustness checks, alternative approaches have been followed. Firstly, imputation has been done by assuming that firms which do not innovate also do not use any sources of knowledge. A possible limitation of such approach is that it could not be appropriate for inputs other than R&D as the acquisition of machinery, equipments, software or other external sources of knowledge: it seems a very strong and unrealistic assumption to consider that these inputs are absent in firms which have not engaged into an innovative project.

Secondly, another way to deal with this problem is to estimate the model only on the sub-sample of firms which have an innovative project, and hence are asked the questions about R&D and other inputs. A first drawback of this approach is that it introduces a quite severe selection bias. Secondly, the large majority of firms engaged into an innovative project successfully innovate. This implies that there are very few 0s in the innovation function and in this case it cannot be estimated in a straightforward way.

Many non reported estimates (available upon request) have been performed. First, we adopt the quite common practice of inputting zero values for the innovation inputs as long as the firm declared that it was not engaged into an innovative project. While the estimated coefficient of internal R&D is comparable both in terms of magnitude and significance compared with that obtained in our preferred estimates in Table 2, the estimated coefficient of other inputs as external R&D, machinery, equipment, software become in many cases non significant and this could be due to the underling assumption of such imputation.

Secondly, the innovation function has been estimated on the subsample of firms engaged into an innovative activity using both the usual probit and the scobit methods. In both cases we obtain non satisfactory results: almost all innovation inputs do not significantly affect the propensity to innovate in the probit regression while in the estimation of the scobit regression, in many cases the Newton-Raphson algorithm does not converge. When it does, we find a non significant effect of external R&D and other external sources of knowledge.

This analysis suggest that among such different strategies to deal with the filter of the CIS questionnaire, that adopted in this paper provides the more credible and satisfactory results.

2.1 The choice of instruments and sensitivity analysis

The choice of the instruments is crucial because is related with the size of the finite sample bias and with the precision of the IV estimates (Bound et al. 1995; Staiger and Stock 1997; Hann and Hausmann 2003). With respect to the bias of the IV-2SLS estimator, it has been shown that it decreases as the sample size increases, as the multiple correlation between the instruments and the endogenous explanatory variable increases, and as the degree of over-identification (the number of instruments) decreases⁷. Consequently, it is important to avoid the use of instruments weakly correlated with the endogenous variable. Moreover the use of “weak instruments” is likely to produces estimates with large standard errors⁸.

These results and the relative small size of our sample suggest to use a small number of “strong instruments”. The instruments used in our preferred estimates (Table 4) have been chosen as follows: first, we run a OLS regression d on ${\ln}\, x_{1},\ldots,\ln\, x_{m},z_{1},\ldots,z_{l}$ and use the backward selection algorithm to exclude the weak instruments across the z ₁,…, z _l . We choose a quite high level of significance, that is 0, 001.

It can be interesting to check the sensitivity of the estimated coefficient of innovation on the choice of the instruments’ set. Various estimates are provided in Table 8 which is organised as follows. Regression (i) reproduces our preferred estimates which have been carried out by selecting the instruments as described above. Regressions (ii) to (ix) use other set of instruments ranging from a minimum of 3 to a maximum of 7 instruments ⁹.

Table 8 The production function (dependent variable: ln(value added), Eq. 1) −2SLS estimates: sensitivity of the estimated effect of innovation on the choice of the instruments’ set

These results shows a very robust picture: in all cases both the magnitude of the estimated coefficient and the significance level are stable across the different specifications. Moreover, in all cases the instruments are found to be orthogonal to the error process.

In particular the coefficients of innovation tout court, product innovation and technological innovation are very stable, always significant and are estimated at about 0.2, 0.25 and 0.25, respectively. It is also worth noting that even if the first-step results indicate high correlated instruments for the three kinds of innovation, such correlation decreases passing from a broad definition of innovation (innovation tout court) to a narrow definition of innovation.

At the opposite, the coefficients of process innovation and non technological innovation are non significant in almost all specifications. The parameter associated to non technological innovation presents a high variability ranging from 0.32 to 0.55.