V. Comparing the Contribution of R&D to Productivity Gains of International Manufacturing Firms

R&D to Productivity Gains of International Manufacturing Firms

SUMMARY

The main purpose of this chapter is to present a set of estimates of the contribution of R&D to productivity performance of international manufacturing firms. The main dataset used to perform this analysis is the balanced panel of 625 firms over the period 1987 to 1994. To begin with, the main outlines of the production function framework as well as the econometric methodology used to estimate the R&D contribution are given The stress is put on the main issues and their possible remedies that this framework raises. A review of the main empirical findings of the literature follows. The empirical analysis starts by presenting a first set of estimates which replicate and update previous studies on R&D and productivity at the firm level. Then, in order to better appreciate to what extent the contribution of firms’

R&D vary across industries, countries and over time, additional estimates based on

different ‘cuts’ of the dataset are discussed. Finally, further results are provided by

applying recently developed GMM estimation methods. These methods deal with

econometric issues such as correlated unobserved firm specific effects and

predetermined explanatory variables which arise in attempting to estimate

production function from panel data.

“Economies of scale, industrial organization, and quality of the labor force are relevant variables which can account for some of the unexplained increase in net national product.

However it is usually agreed that the greater portion of this increase has resulted from technological changes causing a substantial increase in productivity.”

Jora R Minasian (1962: p.94)

5.1. INTRODUCTION

Since the pioneering work of Robert Solow (1957), many econometric studies have given prominence to the contribution of Research and Development (R&D) activities to the economic performances of firms

. It is a well known fact that firms of the most industrialized countries invest substantial amounts in such activities and a natural question arises then, namely what are the success (or the unsuccess) of such investments in terms of economic performances or in other words, do the returns of such activities justify the investments involved. Indeed, R&D activities add to the firms’ existing stock of accumulated knowledge.

This R&D capital stock aims at improving the quality or at reducing the average production costs of existing goods and services or simply at extending the range of intermediate inputs or final goods available to other economic agents. As a consequence R&D activities are likely to affect the firms’ productivity performance. Furthermore, in an ever increasingly global economy, there seem to be different attitudes towards the effort in R&D undertaken by firms of different nations. Many factors may explain these different behaviors towards R&D and a question of interest is then to what extent does the contribution of R&D to productivity of manufacturing firms differ across countries. However, the geographic localization is not the only determinant that may explain differences in the levels of R&D efforts carried out by firms.

Indeed, industry and opportunity effects as well as firm specific characteristics may also induce different behaviors in the decision to devote resources towards R&D activities.

The main purpose of this chapter is to answer these questions by presenting a set of estimates of the contribution of R&D to productivity performance of international manufacturing firms. The main dataset used to perform this analysis is the balanced sample of 625 firms over the period 1987 to 1994 whose construction and main characteristics have been presented in Chapter 3. Yet, the empirical findings that we can get in carrying out such an exercise are likely to be affected by several issues such as the measurement of variables

73 This chapter updates prior work on the R&D-productivity relationship (Cincera, 1995).

74 For a review, see Griliches (1973, 1979, 1994, 1995), Mairesse and Mohnen (1990, 1995), Mairesse and Sassenou (1991) or Capron (1993).

entering the R&D-productivity model, the specification of this relationship as well as issues specific to the estimation of this model through econometric panel data methods. Hence, in our attempt to learn more about the productivity performance of large international manufacturing firms, it is necessary to know what are the issues we have to deal with and possibly what are their remedies and implications in terms of the economic interpretation of the results obtained.

The purpose of Section 5.2 is to summarize the R&D-productivity model (sometimes referred to as the residual productivity methodology) as well as the panel data econometric framework to be implemented in order to estimate the R&D contribution to productivity at the micro level. A specific attention is paid to variable measurement and estimation issues that are particular to this kind of application as well as on the available means to circumvent these issues. Then, in Section 5.3, the most prominent empirical findings of some selected econometric studies on the R&D-productivity relationship at the firm level are summarized.

The main results obtained in this chapter are discussed through Section 5.4. To begin with, a first set of results regarding the R&D contribution are exposed in Section 5.4.2. These estimates replicate the results found in previous studies and hinge on different assumptions regarding:

(i) the dimension of data and the underlying panel data model retained, i.e. cross sectional dimension of data versus temporal one;

(ii) the composition of the datasets on which the estimates are based, i.e. the large sample (S2445) versus the balanced one (S625);

(iii) the estimation of an elasticity or a rate of return;

(iv) the rates of obsolescence of the constructed R&D capital;

(v) the construction of the physical capital;

(vi) the retained price deflator of output;

(vii) the inclusion or not of the rate of exogenous technological change;

(viii) the restrictions imposed on returns to scale; and

(ix) the inclusion of dummy variables that capture industry and geographic specific effects.

Section 5.4.3 examines the extent to which the contribution of R&D to productivity

increases differs across firms of different industries. Thanks to the international dimension of

the dataset, a comparison of the relative impact of R&D activities is also carried out by

estimating separate sets of regressions for firms of different geographic areas. The question of

whether the returns to R&D are related to firm size and R&D intensity is also explored. Finally

the last point in Section 5.4.3. analyses how the R&D contribution has evolved over the period

1987-1994 by running regressions on two sub-periods and on yearly cross-sections of data.

The last empirical section investigates the simultaneity issue associated with the decision in investing in R&D, i.e. does R&D cause output or conversely? First, estimates based on alternative dating of both physical and R&D capitals are discussed. Second, further results are provided by applying recently developed GMM estimation methods. These methods deal with econometric issues such as correlated unobserved firm specific effects and the predeterminency of explanatory variables which arise in attempting to estimate production function from panel data.

5.2. PRODUCTION RESIDUAL METHODOLOGY AND ECONOMETRIC FRAMEWORK

This section aims at reviewing the production function framework

as well as the econometric methodology to be implemented for estimating the contribution of R&D. The stress is put on the main difficulties and limitations of this methodology

A review of the most prominent empirical findings reported in the literature of R&D and productivity will follow.

5.2.1. Production function framework

Most of the econometric studies that have assessed the contribution of R&D to productivity adopt a general version of the Cobb-Douglas production function. This function includes besides the traditional inputs, the firm knowledge capital:

Y

= λ tL C K e

(5.1)

where: Y is output (value added or net sales

);

i and t denote firms and time periods (years) respectively;

e is the exponential function;

t is a time trend assumed to represent the rate of disembodied or autonomous

75 It should be noted that, besides this ‘primal’ approach, another way to model the contribution of R&D has been followed in the literature. This so called ‘dual’ approach which usually rests on a representation of technology by a cost function and from which a system of factor demand equations is then estimated, has not been considered in the present analysis. Among others, Mohnen, Nadiri and Prucha (1986), Bernstein and Nadiri (1988, 1991), Bernstein (1989), Mohnen and Lépine (1991) and Bernstein and Mohnen (1995) have implemented the dual approach at the meso-economic level. The rates of returns from R&D activities obtained in these studies range between 5% and 55%. See also Mairesse and Mohnen (1995) and Van Pottelsberghe (1998) for a review.

76 For a more formal and complete discussion, see Defay (1973), Griliches (1979, 1995), Capron (1993), Mairesse and Mohnen (1995), Hall and Mairesse (1995) or Mairesse and Hall (1996).

77 If sales is the left-hand variable, then materials should be added in the list of inputs. However, this last variable is not always available at the firm level. Value added is sometimes proxied by gross output, i.e. output less changes in inventories of finished goods.

technological change;

L and C are the traditional inputs, i.e. labor and physical capital;

K is the knowledge capital;

ε is a multiplicative error term;

α, β, γ are the parameters of interest, i.e. the elasticities of output with respect to each of the inputs.

Usually, equation 5.1 is taken in logarithm to implement the estimation of α, β and γ.

This leads to the following linear regression model:

y

= + λ α t l

+ β c

+ γ k

+ ε

(5.2)

where lower case letters denote logarithms of variables.

Regarding the explicit functional form of the production function many studies are based on the Cobb Douglas type assuming implicitly a unitary elasticity of substitution between factors. According to Griliches (1979), the choice of the functional form is not very crucial to the extent that no specific interaction between factors is present. For instance, Levy and Terleckij (1989) in their study of Information Technology industries, have estimated the more general CES production function and do not reject the hypothesis of an elasticity of substitution equal to one. Defay (1973) reaches a similar conclusion.

In the way equation 5.2 is specified, returns to scale with respect to the three inputs can be assumed constant or not. There are constant returns to scale if the sum of factors elasticities is equal to one. In order to measure explicitly this assumption, equation 5.2 is generally re- written by substracting labor from both sides of this equation:

( ^y

− ^l

) = + + − ^a λ ^t ( µ 1 ) ^l

+ β ( ^c

− ^l

) ( + γ ^k

− ^l

) + ε

(5.3)

where: µ=α+β+γ represents the coefficient of returns to scale.

Another important difficulty raised by such a specification is related to the construction of the knowledge capital for the firm. Actually, the perpetual inventory method originally proposed by Griliches (1979) is the most commonly used method for constructing the firm knowledge capital. As it has be seen in Chapter 3, this method assumes that the current state of knowledge is a result of present and past R&D expenditures discounted by a certain rate of depreciation.

As seen in Chapter 3, this formulation suffers from some drawbacks. First, the

magnitude of the depreciation rate is unknown. Second, since the available history of R&D is

usually not very long, we need a method to construct the initial knowledge stock. Third, it is not clear whether the assumed distribution of the R&D effects is the most appropriate.

In order to get around the issues associated with the construction of the R&D stock of knowledge, an alternative specification to equation 5.2 is sometimes used. This approach, suggested by Griliches (1973) and Terleckij (1974), directly estimates the rate of return to R&D instead of its elasticity. This approach departs from an alternative formulation of equation 5.2 where levels of the variables are replaced by growth rates. More specifically, we have from equation 5.1:

γ ∂

∂ ρ

= Y =

K K Y

K Y

it it

it it

it it

(5.4)

where ρ is the rate of return to R&D.

Since the growth rate of variables in equation 5.1 can be approximated by the first difference of their logarithms, we have:

( )

γ∆ γ ρ ρ ρ δ

ρ δ

ρ

k K

K

K Y

K K

K Y

K R K

Y

R K

Y

R

it it

Y

it

it it

it it

it it

it it it

it

it it

it

it it

= = = = − + −

= − ≈

− − −

∆ ∆ ∆ 1

(5.5) provided that the rate of depreciation of the R&D capital is close to zero. Hence, equation 5.2 expressed in growth rates, can be re-written as:

∆y t l c R

it it it

Y

it

= λ∆ α∆ + + β∆ + ρ + ε

(5.6)

As pointed out by Capron (1993), this alternative approach turns out to be more

consistent with the optimal R&D choice behavior of firms compared to the elasticity approach

that assumes a common elasticity of output with respect to R&D capital when the relationship

is estimated across firms. Indeed, to the extent that the production technology is specific to

each firm, firms will use different factor shares and if inputs are used at their competitive

equilibrium levels, firms are unlikely to have the same output elasticities. Yet, the hypothesis of

the absence of depreciation for the R&D capital stock is more questionable. Quoting Capron

(1993: p.66), “ At the firm level, the marketable knowledge ensuing from its R&D investment

decreases over time because upgraded products and processes reduce its market valuation and

because the privately acquired knowledge leaks out to competitors. ”

Two other well known issues encountered when estimating the contribution of R&D remain to be mentioned. The first problem is the so called ‘double counting’ of R&D. This double counting arises since the conventional inputs generally already include the R&D-labor and R&D-capital components of R&D expenditures. As shown by Schankerman (1981) and Mairesse and Hall (1996) for instance, this double counting reflects itself in downward estimates of R&D elasticities and rates of returns. As a consequence, when the physical and labor are not cleaned from their R&D components, the rate of return to R&D has to be interpreted as an excess rate

. The second issue is related to the way current and past values of R&D investments have to be deflated when measuring the R&D capital. Some authors have paid attention to this issue by constructing ‘counpound’ and ‘two digit level’

price indexes, but there seems to be no substantial differences in the results according to whether these price indexes or the GNP deflator are used.

Finally, besides these issues specific to the R&D variable, there are also problems when measuring output and the other inputs entering the production function. Regarding the former variable, one of the most acute issues, at least at the micro level, is the way it has to be deflated. The first important drawback is that price deflators are usually not available at the firm level. Instead more aggregate price indexes are used, in general at the industry two digit level which raises several problems for industries characterized by imperfect competition

or for large firms which are multi-products and have subsidiaries in many countries. For instance, if a firm manufactures two products, one in country A and the other in country B, which price deflator of which country has to be used if we only observe total sales? Moreover, if these products are in two different two digit level industries, e.g. a drug and a chemical product, which price deflator of which industry do we have to retain? The second shortcoming is that such price deflators do not incorporate output quality changes and as a result underestimate the ‘real’ output. The computer industry is a well known example. Mairesse and Hall (1996), for their US sample use an hedonic price index for computers which captures the substantial quality improvement experienced by this industry during the last two decades. They report an estimated R&D capital elasticity of .25 which is much higher than the corresponding result of .04 obtained by deflating output by the more conventional GNP deflator. However, as

78 Quoting Mairesse and Hall (1996: p.5), “Conceptually, the value added, labor, and capital measures used to estimate [the productivity equation] should be purged of the contribution of R&D materials, physical capital used in R&D laboratories, and R&D personnel, since these inputs do not produce current output, but are used to increase the stock of R&D capital. If this is not done, the cross section estimates [...] will not necessarily be incorrect, but the measured R&D coefficient will be some kind of ‘excess’ elasticity of output to R&D rather than a total elasticity, i.e. the incremental productivity of R&D rather than a total elasticity.”

79 Bernstein (1986) has constructed for Canada a Divisia price index that incorporates the prices of different components of R&D, while Mansfield, Romeo and Switzer (1983) have considered a Laspeyre price index.

Mairesse and Hall (1996) compare the GNP deflator with industry two-digit level ones.

80 Quoting Griliches (1995: p.74), “if the producer does not have a complete monopoly power over his invention, the price he receives will not reflect all the potential social benefit since part of it will be passed on to the consumers in the form of lower prices per equivalent quality or performance unit.”

Mairesse and Mohnen (1995) emphasizes, with panel data, quality differences can be captured by time and sector dummies, even in the absence of good prices. However, there remain the inter-firm differences, which are not captured by these dummies. The R&D estimates are thus biased but only to the extent that sector prices or dummies do not fully capture the quality differences and the latter are correlated with the explanatory variables.

Regarding the measurement of the traditional inputs, correction for quality differences in labor and physical capital should be allowed for as well. Turning to the previous computer example, the contribution of such devices to productivity gains of firms using them as inputs of the production function will be underestimated if quality changes are not taken into account.

Furthermore, according to Griliches and Mairesse (1984), as long as the inputs are not corrected by the maximal production capacity rates, variations of these inputs affect the productivity’s measurement. Here also, if we assume that these rates of capacity utilization are more or less similar among firms within a given industry and for a given time period, then such business cycles effects should be attenuated by including appropriate industry and time dummies.

5.2.2. Econometric estimation methods

Among the different approaches for estimating the relationship between R&D and productivity, many studies have adopted panel data econometric methods

. In such a setting, the typical dataset contains observations on a cross section of firms over several time periods.

Most often, there is a large number of cross-sectional units, e.g. 100, 500,... firms and only a few time periods, e.g. 2, 5 or 10 years. Consequently, the cross-sectional variation is in general much larger than the temporal one. Similarly, the main source of heterogeneity is found across firms. This double dimension confers to panel data several advantages with respect to purely cross section or time series data. Among the main advantages, we can mention more informative data, more variability, less collinearity among the variables, more degrees of freedom and more efficiency (Baltagi, 1995). However, the main benefit is probably the greater flexibility in the modelling of behavioral differences across individual units.

A useful way to introduce these alternative modelling possibilities, is to have a closer look at the structure of the error term of equation 5.2. More specifically, this error term can be decomposed as follows:

81 For an introduction to the econometrics of panel data, see Hsiao (1986), Baltagi (1995), or Matyas and Sevestre (1996) for instance.

uit = α i + λ t + ε it (5.7) where: α i is the cross sectional unit or firm specific effect, e.g. the ability of the R&D

personnel to find new inventions

;

λ t is a time-period specific effect, e.g. technological change, exchange rate variation;

ε it summarizes the effects of left-out factors, approximation errors and other disturbances.

The less restrictive model is the one that assumes the αi and the λt to be constant across all units and all time periods, respectively. In this case, the estimates are based on the levels of the variables and by pooling the cross-section and time-series observations composing the panel dataset. The estimates obtained from this model are usually referred as to the (level-) total estimates. Alternatively, a regression of equation 5.2 can be run for each time-period separately ((level-) cross section estimates) or by averaging the observations of each cross- sectional units over all time-periods (between estimates). This last model is equivalent to remove the last two terms in equation 5.7. In the models discussed so far, the stress is put on the cross sectional dimension of the data. The next three models focus on the time-series dimension of data. The first-differences (total) estimates are based on annual growth rates of the variables while taking deviations from individual means of equation 5.2 leads to the within estimates

. The within model corresponds to growth rates of the variables in equation 5.2 when they are taken in logarithms. Moreover, the within transformation is equivalent to the introduction of individual dummies in the first model, i.e. total estimates. More precisely, these individual dummies are equivalent to the αi provided that we can assume that they are fixed over time. These parameters are sometimes referred to as firms’ specific unobserved fixed effects. They can be viewed as parametric shifts of the productivity model that reflect differences between firms. Formally, by ‘first differencing’ equation 5.7, we have:

∆ uit= ∆λ t + ∆ε it (5.8)

while applying the within transformation leads to:

82 R&D opportunity or managerial skills may also be mentioned. Quoting Salter (1969: p.88-89), “Differences in the personal skill, effort, intelligence and co-operation of labour may alone lead to substantial inter-plant variations in productivity. Equally, if not more, important are variations in the efficiency of management which are not reflected in the managers’ salaries; an efficiently managed firm employing outmoded capital equipment may achieve lower operating costs than a poorly managed firm using modern equipment. Other special advantages, such as favourable location, access to ancillary services, trade goodwill, ect., may also contribute to inter-plant differences in operating costs and productivity. Barriers to the diffusion of knowledge, especially the patent system, are also relevant in this context. Some plants may employ outmoded methods, not because replacement is unprofitable, but simply because patent restriction prevent the use of the best methods.

Other restrictions, such as imperfect channels for the diffusion of technical knowledge, ignorance and inertia, may have the same effects.”

83 The growth rates specification can be pooled over several time periods (total estimates), this period can be averaged (between estimates) and finally a ‘long’ growth rate between the last and the first period can be performed (long difference estimates).

uit-ui.= λ t- λ t. + ε it- ε i. (5.9) It can be observed that these last two models eliminate the α i provided that they are fixed over time. Another point of view is to consider that these firms’ specific effects are not fixed but randomly distributed across cross-sectional units. In such a setting, we speak about random effectxs. One advantage of random effects over fixed ones, is the higher efficiency of the former estimates. However, the consistency of the random estimates relies on the assumption that the individual effects and the regressors are not correlated. We will see below that such an assumption is rarely verified in the context of the R&D-productivity relationship.

Before discussing further these models, two statistical tests should be presented. The first one is based on a comparison of the total and within estimates and tests the hypothesis of whether the α i are all equal with an F-test. The second test has been derived by Hausman (1978). It allows one to accept or to reject the null hypothesis that the individual effects are uncorrelated with the regressors.

From the analysis of panel data, we know that the estimates obtained from these different models will be consistent and similar provided that equation 5.2 is correctly specified, that is, if all the relevant explanatory variables are included in equation 5.2. In practice, however, there are very often substantial discrepancies in the estimates across these models, in particular between the models that favor the cross-sectional dimension of data and the ones that exploit the temporal dimension. Several arguments can be put forward to explain these important disparities of the results when we move from one model to the other.

A first argument is related to the specification of the R&D-productivity model. Indeed, if some important explanatory variables are missing from the productivity equation, and to the extent that these omitted variables are correlated with the regressors at hand, then omitted variable biases are likely to prevail in the estimates. For instance in the context of the R&D- productivity relationship, the level of technological opportunity available to a research laboratory of a given firm, the ability of its engineers to find a new product, or the efforts made by the R&D personnel to succeed are typical unobserved variable of firms. These unobservables are likely to be ‘transmitted’ to the R&D decision since firms facing higher opportunities or abilities will generally invest more in research activities. To put it differently, we are in the presence of a (positive) correlation between these unobservables and the R&D variable.

Yet, in the context of panel data, it is possible to get around this issue by using ad-hoc

transformations of data. Indeed, if we can make the assumption that the left-out variables in the

productivity equation are largely fixed over time, that is to say, they change only in the long

run and not in the short time span characterizing the length of the panel dataset

, then it can be considered that the α i of equation 5.7 capture these unobserved effects. In turn, these firms’

specific unobserved effects can be ‘eliminated’ by first-differencing the productivity equation (as in equation 5.8) or by applying the within transformation (as in equation 5.9). Still, it should be noted that the estimates based on these transformed models will be consistent if and only if some conditions regarding the exogeneity of the regressors are satisfied. For the within estimates, strong exogeneity of the regressors is required, i.e.

E[xis ε it]=0, ∀s = 1,...,T and ∀t = 1,...,T (5.10)

since subtracting individual means from equation 5.7 contaminates the ε it’s with the disturbances from the other years, ε i1,..., ε iT. Unfortunately, the strict exogeneity condition is another hypothesis which is hard to maintain in the R&D-productivity framework as it will be discussed below. The consistency of the first difference estimates relies on a less restrictive hypothesis to the extent that the regressors need only to be predetermined, i.e.

E[xis ε it]=0, ∀s = 1,...,t-1 and ∀t = 1,...,T (5.11)

since first differencing equation 5.7 introduces ε it-1 in equation 5.7. Hence, the first difference transformation may be preferred to the within one to the extent that the exogeneity assumption is less restrictive for the former transformation.

However, the presence of random measurement errors in the variables renders the first difference estimates generally more biased towards zero than the within correction (Griliches and Hausman (1986)). More generally, the estimates based on the first difference and within transformations give poorer results than the ones in the cross sectional dimension of data.

Among the main causes explaining these less satisfactory results, Mairesse and Sassenou (1991) mention, besides the exogeneity hypothesis and the measurement issue, the high collinearity of R&D capital with time, inadequate specification of lags in the effects of R&D capital and the omission of variables reflecting short-term adjustments. These outcomes constitute the second main cause of the differences observed in the two kinds of estimates.

To sum up the problems discussed so far, in panel data models, we have the choice between econometric models in the cross sectional dimension of data and models in the time series dimension. If some important explanatory variables are not included in the productivity

84 Quoting Jaffe (1986: p.986), “If technological opportunity affects profitability, then we would expect firms to move to the more profitable [technological] positions. However, changes in the technological position of firms can be brought only slowly. Expertise in various areas is not easily acquired, and goodwill and reputation in product markets represent sunk costs that make jumping costly.”

equation, and to the extent that these omitted variables are correlated with the ones at hand, then omission biases are likely to be present in the estimates based on cross sectional models.

In order to get around this problem, we can assume that these unobserved variables are picked up by firms’ specific effects, which in turn can be removed by applying the first difference or the within transformations. However, these models require the predeterminency or even the strong exogeneity of regressors for the within estimates to be satisfied. Moreover, various issues, in particular random measurement errors in the variables, may exacerbate the omitted variable bias by deteriorating the estimates we can get by applying these transformations.

The strong exogeneity of the regressors is related to the simultaneity of the decision process regarding employment or R&D and production. This last issue is a third main cause of the high disparities between the two kinds of estimates. It refers to the question of whether R&D, for instance, depends on past, current or future values of output, i.e. expectations of the dependent variables, or conversely. In other words, do we measure the effect of R&D on productivity or conversely, are we estimating for a large part the counterpart of the influence of output on R&D?

Among the different solutions proposed to circumvent this simultaneity problem, semi- reduced forms of equation 5.2 can be estimated (Griliches and Mairesse, 1984). Mairesse and Hall (1995) consider beginning of years capital stocks rather than end of years to attenuate the possible simultaneity biases. Griliches, Hall and Pakes (1991) ‘enrich’ the R&D-productivity model by considering additional information on the investment policy adopted by the firm, the number of patents it has received or its stock market value. In particular, the expectation mechanisms of output is formalized by relating this variable to the stock market value of the firm. An alternative approach that allows for all these issues to be present, i.e. correlated fixed effects, measurement errors and simultaneity, is the one implemented in Chapter 4. This more general approach relies on a General Method of Moments (GMM) estimator in the context of linear panel data econometric models

. As a starting point, the GMM framework assumes the presence of correlated fixed effects with regressors in the productivity equation (so that this equation has to be first-differenced

) and by assuming that only lag 2 or higher values, i.e. lag 3, 4,..., of regressors are available as instruments (because later values are correlated with the error term)

. These assumptions imply the following set of orthogonality conditions:

85 See Mairesse and Hall (1996) for a general description of this methodology (which is based on that of Arellano and Bond (1991), Keane and Runkle (1992), and Schmidt, Ahn and Wyhowski (1992)) and its empirical implementation. A similar framework is used by Klette (1994).

86 The first difference transformation is ‘preferable’ to the within one because of the strict exogeneity of regressors required by the latter.

87 Mairesse and Hall (1996) are more restrictive, since their maintained model is that only lag 3 and higher values of the Xit are uncorrelated with the εit. The authors (see footnote 15, p.12 in their paper) ground this choice on prior experience with firm data. They add that this choice errs on the side of caution, and may not be strictly necessary depending on the dataset at hand.

E[ Xis

it ]=0 where i=1,...,N; s=1,...,t-2 and t=3,...,T. (5.12)

These moment conditions are then estimated by means of the GMM. Relaxing the assumption of lag 2 or higher values of regressors as valid instruments, implies additional moment conditions. Testing the validity of these additional conditions allows to determine if the regressors are lag 2, lag 1, weakly or strongly exogenous. Furthermore, if the fixed effects are not correlated with the regressors, then it is appropriate to estimate the productivity equation in level. Here again, considering this equation in levels rather than in first-differences implies additional moment conditions whose validity can be tested in order to answer the question of correlated fixed effects.

It should be noted that though this GMM framework seems quite attractive in terms of its modelling possibilities and the weak distributional assumptions on which it relies, it nevertheless rests on an instrumental variable approach and according to Griliches and Mairesse (1995), the past levels as instruments for current growth rates of regressors are likely to be quite poor and possess little resolving power.

5.3. THE R&D CONTRIBUTION TO FIRMS’ PRODUCTIVITY: REVIEW OF SOME SELECTED STUDIES

This section reviews some of the main econometric studies that have investigated the R&D-productivity relationship at the micro level. This survey will give a better idea of the order of magnitude of the R&D contribution to the productivity performances of firms. It will also illustrate some of the methodological and technical issues discussed in the previous section as well as their implication in terms of the conclusions drawn from the estimates.

First of all, it is worth saying that these studies are far from being fully comparable despite the fact that they are based on the same standard framework. This lack of comparability can be explained by various reasons.

First, the data samples differ to a great extent among studies, though they generally

concerns firms operating in the manufacturing sector. Among the main differences

characterizing these data samples, the period under investigation, the sample size, the country

of origin of firms can be mentioned. For instance, the most aged study in Table 5.1 is the study

of Minasian who examined the R&D-productivity relationship of firms over the period 1947 to

1957. At the opposite the most recent period analyzed is found in the study of Hall and Mairesse (1995) which covers the 1981-1989 period. Obviously, it can be expected that the contribution of R&D to productivity has changed over time. Hence, one should not be astonished by changes from one period to the other in the estimates. Regarding the size of data samples, the study of Minasian (1969) consists of 17 firms while Hall and Mairesse (1995) carry out their analysis on the basis of a sample of 1232 French firms. Here also, it can be expected that the estimated returns of R&D will depend on the number of firms to the extent that the R&D contribution changes with firm size. For instance, the larger the sample is, the more it will include small firms so that the average firm’s size will tend to be smaller as the sample’s size increases.

Not only may R&D intensities vary over time but also across countries and industries.

As a consequence, the estimated contribution of R&D to productivity may be different for firms of different countries. It is interesting to note that a large majority of studies concentrate on the R&D-productivity relationship in the United States. Indeed, it appears from Table 5.1 that 15 studies out of 25 ground their analysis on US firms. This is especially the case for studies at the beginning of the period. The other two countries for which several studies have been performed are Japan and France. Belgium, the Netherlands and Germany are also represented, but only by one study. It is also worth mentioning that no study has been made at least at the micro level and as far as we know, for Italy and the United Kingdom which are two other countries in which a substantial number of firms are engaged in R&D activities. One reason for that situation is the difficulty of having access to the relevant information at the firm level. Besides these differences, some studies have also restricted the scope of their analysis to firms belonging to a same industrial sector like the sector of chemicals in the study of Mansfield (1980).

The second explanation for the differences in the findings of these studies rests on the way some variables entering the production function have been constructed as well as the modelling of this function itself. All the studies reported in Tables 5.1 to 5.3 hinge on alternative specifications of the primal approach and/or econometric panel data models discussed in Section 5.2

, traditional inputs and R&D capital of firms. To ease the comparability of the studies reported in these tables, they have all been ranked according to different criteria:

• variables entering the production function taken in level or in first differences;

• estimation of the R&D elasticity or the rate of return to R&D with respect to output;

88 i.e. the estimation of a Cobb-Douglas production function which relates the logarithm of output to the logarithm of traditional inputs and firm’s own R&D capital.

• sales or value added is considered as output of the production function.

Besides these different specifications, other differences have to be mentioned. First, some studies have included a time trend which is assumed to pick up the autonomous rate of technological change while other studies introduce additional determinants of productivity such as variables that reflect business cycle effects or market concentrations. Very often, these additional determinants are proxied by industry dummies. Second, different kinds of price indices have been used to deflate the variables entering the production function. Third, some studies have corrected the double counting of R&D in traditional inputs.

Finally, the third major source of differences in the estimates is related to the econometric panel data model used to perform the estimates. Hence, besides the models that estimate the level of the variables entering the productivity model, alternative models that favor the temporal dimension of data are also estimated. These alternative transformations of the production function allow for a fixed effect to be present. Here again, the main findings reported in Table 5.1 have been classified according to the econometric models implemented.

Tables 5.1 to 5.3 indicate that for a majority of studies, it is the output elasticity of the R&D stock which has been estimated while a less important number of studies have focused their attention on the private rate of return to R&D. Regarding the elasticity of R&D stock approach, half of studies have considered a level specification of the production function. The second half has specified this function in the temporal dimension.

(i) elasticity of R&D: level dimension estimates

The studies reported in Table 5.1 have assessed the elasticity of R&D stock on output

on the basis of a level specification of the production function. The first four studies are based

on firm’s cross sectional data. In the first study (Schankerman, 1981), the estimates are

performed for different industry sectors. The estimates that are statistically significant range

from .034 (Electric equipment) to .292 (Aircraft industry). For each industry, additional

estimates that try to correct for the double counting issue, are also presented. According to

Schankerman (1981), it is important to correct traditional inputs, i.e. labor and physical capital,

for double counting of R&D inputs. When such a correction is not operated, downward biases

are likely to arise since a part of the traditional inputs is used to increase the stock of R&D

capital and not to produce the current output. Indeed, his results indicate that such a bias is

present and quite important for some industries such as Aircraft (800%) and Electric

equipment (600%). For other industries such as Chemical and oil and Motor vehicles, the

downward bias is still present but less important: respectively 50% and 30%. It is interesting to note that the highest estimated elasticities of R&D have to be found for the two most R&D intensive industries which are Aircraft (R&D elasticity of .29) and Electric equipment (.23).

Moreover, these industries also present the largest downward biases arising from the double counting issue.

In Griliches (1980, study 2), the estimated elasticity of R&D with respect to output is equal to .07 which is less than the corresponding result of .12 reported in his 1986’s study (study 3). This difference can be explained by a lower number of firms in the latter study and hence, the higher R&D intensiveness characterizing the sample. Ceteris paribus, it could also be argued that the contribution of R&D has been higher during the mid 70’s in comparison to the early 60’s.

However, study 3 also reports estimates with industry dummies as additional explanatory variables of the production function. In this case, the estimated R&D elasticity is lower with a value of .09. This different result gives clue that the omission of industry specific variables, whose effects can be picked up by industry dummies, may lead to biased estimates.

Actually, the interpretation of these sectorial dummies is ambiguous. Quoting Mairesse and Mohnen (1995: p.37), “ On the one hand, the indicators may correct the estimates for the bias resulting from the erroneous omission from the production function of structural variables strongly correlated to the sectorial characteristics. On the other hand, their presence may itself be a source of distortion to the extent that they reflect in part the return to research resulting from technological opportunities. The latter are probably essential to explain the greater tendency to carry out research in certain sectors. Thus, scientific sectors benefit from a more solid and broader knowledge base, on which it may be easier to devise a research program and achieve profitable innovations ” .

Sassenou (1988, study 4) reaches a similar conclusion in his analysis of Japanese firms.

Indeed, the elasticities of R&D reported in his study are once more lower when industry dummies are included. Another interesting finding reported in Sassenou’s study is the somewhat larger elasticity of R&D for Scientific firms (.16) against other firms (.10).

Consequently, the contribution of the stock of R&D capital is more important for more R&D intensive firms. Finally, comparing the estimates of studies 3 and 4, it follows that the contribution of R&D is quite similar for US and Japanese firms.

The next series of studies (study 5 to 11) exploits both the cross-section and temporal

dimension of data. One advantage of such a panel structure of data is to increase substantially

the degrees of freedom. One of the first studies that has examined the R&D contribution to

output performance is the one of Minasian (1969, study 5). His estimated contribution of R&D

is performed on a small sample of 17 US Chemicals firms. Minasian estimated an elasticity of

.11 which is very close to the one obtained by Schankerman (1981, study 1) for firms performing in the same industry sector during the early 60’s.

Table 5.1

Selected Productivity-R&D studies at the micro level: elasticity of R&D, level dimension^a

# study data specification^b econometric

model

est.^c s.e.^d

1 Schankerman (1981) 63, 110 US firms, Chemicals and Oil VA Cross-section .104 ^(.036)*

" " VA-DC " .159 ^(.035)*

" 63, 187 US firms, Metals & Machinery VA " .018 ^(.022)

" " VA-DC " .099 ^(.021)*

" 63, 101 US firms, Electric equipment VA " .034 ^(.020)**

" " VA-DC " .232 ^(.029)*

" 63, 34 US firms, Motor vehicles VA " .069 ^(.047)

" " VA-DC " .090 ^(.046)*

" 63, 31 US firms, Aircraft VA " .032 ^(.033)

" " VA-DC " .292 (.048)*

" 63, 419 US firms, Miscellaneous VA " .043 ^(.011)*

" " VA-DC " .065 ^(.011)*

2 Griliches (1980) 63, 883 US firms VA Cross-section .069 ^(.009)*

3 Griliches (1986) 72, 491 US firms VA-ID Cross-section .115 ^(.018)*

" " " " .089 ^(.017)*

4 Sassenou (1988) 76, 394 Japanese firms VA Cross-section .10 ^(.01)*

" 76, 112 Japanese firms, Scientific " " .16 ^(.03)*

" " ID " .07 ^(.02)*

5 Minasian (1969) 48-57, 17 US firms, Chemicals VA Total .113 ^(.015)*

6 Griliches-Mairesse (1984) 66-77, 133 US firms Sales Total .054 ^(.011)*

" 66-77, 77 US firms, Scientific Sales " .185 ^(.013)*

7 Cunéo-Mairesse (1984) 72-77, 182 French firms VA Total .203 ^(.007)*

" 72-77, 98 French firms, Scientific VA " .114 ^(.010)*

" " VA-DC " .206 ^(.014)*

" " Sales " .176 ^(.019)*

8 Harhoff (1994) 77-89, 443 German firms Sales Total .15

9 Hall-Mairesse (1995) 80-87, 197 French firms VA Total .180 ^(.009)*

" " VA-DC " .252 ^(.008)*

10 Mairesse-Hall (1996) 81-89, 1073 US firms Sales Total .035 ^(.005)*

" " Sales- I2Ddef " .246 ^(.012)*

" 81-89, 1232 French firms Sales " .090 ^(.006)*

" " Sales- I2Ddef " .093 ^(.006)*

VA " .092 ^(.004)*

" " VA-DC " .165 ^(.004)*

11 Bartelsman & al. (1996) 85-89, 209 Dutch firms VA Total .008 ^(.016)

" " VA-DC " .046 ^(.015)*

" 89-93, 159 Dutch firms VA " .043 (.023)**

" " VA-DC " .099 ^(.022)*

notes: a) adapted and extended from Mairesse and Sassenou (1991) and Mairesse and Mohnen (1995)

b) VA = value-added; DC = correction for the double counting of R&D; ID = inclusion of industry dummies; I2Ddef = Industry 2 digit level deflators

c) R&D estimated coefficient

d) standard error, *(**) = statistically significant at the 5% (10%) level.

Griliches and Mairesse (1984, study 6) investigate 133 US firms. They find an elasticity of R&D with respect to sales of .05 which is higher than the corresponding result of .04 obtained by Mairesse and Hall (1996, study 10). Here also, the larger sample of the latter study and the more recent period on which the analysis is based may explain these different results.

Griliches and Mairesse report results for Scientific firms as well. Here, the estimated R&D

elasticity is of .19 which is higher than the findings reported for other firms of their sample.

This result is not in contradiction with the result obtained by Sassenou (study 4) for similar firms, but in Japan. Conversely, the R&D elasticity reported by Cunéo and Mairesse (1984, study 7) is lower for Scientific French firms (elasticity of .11) than for other ones (.20). One possible explanation of this result may arise from the fact that both US and Japanese firms in the Scientific sector are much less reliant on government funding than is the case for French firms. Since public-funded R&D is in general more fundamental based, it is likely that its contribution to productivity performance of private firms is less important than the returns of private financed R&D

. Cunéo and Mairesse give alternative results for firms in the Scientific sector. They experiment with value added and sales as the output variables and correct for the double counting of R&D inputs. As far as this correction is concerned, here also a downward bias is observed (about 80%). Considering sales instead of value added as the dependent variable leads to a higher elasticity of R&D capital (.18 for sales versus .11 for value added).

Harhoff (1994, study 8) presents estimates of the R&D elasticity for a sample of 443 German firms. The value of .15 reported in his analysis is quite comparable with similar results obtained for French firms and on the whole, his result appears to be higher than the findings performed for US and to some extent for Japanese firms.

Hall and Mairesse (1995, study 9) obtain an estimate of R&D elasticity of .18 for French firms during the 80’s. This result is very close to the one reported by Cunéo and Mairesse (study 7) though the period investigated in this analysis is not the same (1972-1977).

Furthermore, Hall and Mairesse find a higher R&D elasticity when correction for double counting is allowed for.

The study of Mairesse and Hall (1996, study 10) uses two datasets in both the US and France to investigate the R&D-productivity relationship in both countries and to examine to what extent this relationship is similar in France and the US. For France, the reported estimates show, on the one hand, that using sales instead of value added leads to very similar results and, on the other hand, that attempting to correct for double counting of R&D employees raises the R&D elasticity by about 80% which is consistent with previous findings with regard to this issue. Furthermore, the authors explore two alternative price indexes used to deflate the dependent variable. The first one consists of the whole manufacturing sales deflators while the second one is defined as the sales deflator at the industry two digit level. Using the latter deflator tends to increase slightly the R&D elasticity in France and dramatically in the US (about 500%). For the authors, the reason of this difference comes from the way the US

89 Mansfield (1980), Griliches and Lichtenberg (1984), Griliches (1986) and Lichtenberg and Siegel (1991), for instance, have reported a higher (or more significant) rate of return on private than on public R&D.

deflator is constructed. Actually this deflator is based on a hedonic price index for computers and as a consequence while computers became much more powerful and much less expensive during the 80’s, the price index for this industry declined by about 80% during this period.

Since this industry has become increasingly important in manufacturing, the effects of this price decline combined with growing R&D budget and output productivity play an substantial role in explaining the observed differences in terms of R&D elasticity. Finally, the authors investigate the effect of the dating of both physical and R&D capitals. They experiment the effect of considering end year of capitals versus beginning of year. It can be observed that for the level specification of the production function, the results do not change significantly from one dating to the other.

Bartelsman et al. (1996, study 11) provide estimates of the contribution of R&D for Dutch Manufacturing firms. On the whole, the estimated coefficients reported in their analysis are lower than the corresponding ones obtained in other countries. Here also, the results are sensitive to the double counting of R&D inputs: when double counting is corrected, the R&D elasticity increases from .01 to .05 for the period 1985-1989 and from .04 to .10 for the 1989- 1993 period.

(ii) elasticity of R&D: temporal estimates

The series of studies (studies 12-25) reported in Table 5.2 use the temporal dimension of the data. Some studies are based on the growth rates of the variables (total or between first difference estimates or long difference estimates) while other rest on deviations from the means of the variables (within transformation). As a whole, it can be observed that estimates of R&D elasticity in the temporal dimension provide more controversial results than estimates in the level dimension of the data. In fact, in a large number of studies the estimates coefficients appear to be statistically non significant (studies 13, 14, 16, 19, for instance). Quoting Mairesse and Mohnen (1995: p.37), “ The fact that the estimates are lower and more fragile in the temporal dimension can be explained in a number of way. A simple but important reason relates to the collinearity between the physical capital and research capital variables and the time trend reflecting autonomous technical change. Another reason is that measurement errors tend to have a much more serious impact on growth rates than on the levels of the variables. A further factor is no doubt the omission of cyclical variables in the production function, such as the duration of work, the rate of utilization of physical capital, and more generally, the difficulties of providing a satisfactory specification of the lags and the dynamic evolution of the variables. ”

However, in some cases, the estimated elasticities are close for the level and the time

dimensions of the data. Griliches (1980, study 12) investigates the relationship between the

rate of growth of partial productivity and the growth of R&D during the period 1957-1965.

Despite the fact that the dependent variable used in this study is the partial productivity growth (computed as the difference between the estimated rate of growth of total company sales and the product of the rate of growth of total firm employment and the average share of labor in sales), the implied elasticity of R&D stock is about .08 which is quite similar to the elasticity reported in study 2 performed on a cross-section of the same data sample. The next results are given for several industry sectors. The estimated elasticity appears to be somewhat higher for R&D intensive sectors (about .1) and lower (about .4) for the other firms which represent the other ‘half’ of the sample.

Table 5.2

Selected Productivity-R&D studies at the micro level: elasticity of R&D, temporal dimension^a

# study data specification^b econometric

model

est.^c s.e.^d

12 Griliches (1980) 57-65, 883 US firms PFP-ID Between .076 ^(.013)*

" 57-65, 110 US firms, Chemicals &

Petroleum

PFP " .093 ^(.038)*

" 57-65, 187 US firms, Metals &

machinery

" " .102 ^(.022)*

" 57-65, 101 US firms, Electric equipment " " .106 (.030)*

" 57-65, 34 US firms, Motor vehicles " " .126 ^(.070)**

" 57-65, 31 US firms, Aircraft " " .107 ^(.077)

" 57-65, 419 US firms, Other " " .052 ^(.015)*

13 Griliches-Mairesse (1983) 73-78, 343 US +185 French firms Sales Between .02 ^(.03)

14 Sassenou (1988) 73-81, 394 Japanese firms VA Between .02 ^(.02)

15 Mairesse-Hall (1996) 81-89, 1232 French firms Sales F.D. total -.003 ^(.003)

" Sales-I2Ddef " -.003 ^(.003)

" VA " -.005 ^(.003)**

" 81-89, 1073 US firms Sales " .010 ^(.024)

" 81-89, 1073 US firms Sales-I2Ddef " .092 (.026)*

16 Sassenou (1988) 73-81, 394 Japanese firms VA F.D. total .04 (.04)

17 Mairesse-Cunéo (1985) 74, 79, 390 French firms L.D. .02 (.10)

18 Bartelsman & al. (1996) 85-89, 209 Dutch firms VA-DC L.D. .247 (.083)*

" 89-93, 159 Dutch firms " " .104 ^(.080)

19 Minasian (1969) 48-57, 17 US firms, Chemicals VA Within .084 ^(.068)

20 Griliches-Mairesse (1984) 66-77, 133 US firms Sales Within .091 ^(.022)*

" 66-77, 77 US firms, Scientific " " .021 ^(.026)

21 Cunéo-Mairesse (1984) 72-77, 182 French firms, Scientific VA Within .050 ^(.039)

" 72-77, 98 French firms, Scientific " " .144 ^(.054)*

" " VA-DC " .170 ^(.052)*

" " Sales " .028 ^(.043)

22 Sassenou (1988) 73-81, 394 Japanese firms ? Within -.01 ^(.01)

23 Hall-Mairesse (1995) 80-87, 197 French firms VA Within -.001 ^(.036)

" " VA-DC " .069 ^(.035)*

24 Mairesse-Hall (1996) 81-89, 1232 French firms Sales Within .008 ^(.011)

" " Sales I2Ddef " .013 ^(.011)

" " VA " -.016 ^(.013)

" 81-89, 1073 US firms Sales " .041 ^(.011)*

" " Sales I2Ddef " .170 ^(.014)*

25 Harhoff (1994) 77-89, 443 German firms Sales Within .09

notes: a) adapted and extended from Mairesse and Sassenou (1991) and Mairesse and Mohnen (1995)

b) PFP = partial factor productivity; ID = inclusion of industry dummies; VA = value-added; I2Ddef = Industry 2 digit level deflators; DC = correction for the double counting of R&D; F.D. = first difference; L.D. = long difference

c) R&D estimated coefficient

d) standard error, *(**) = statistically significant at the 5% (10%) level

(iii) rates of return to R&D

The last set of studies reported in Table 5.3 (studies 26-41) directly estimate the rate of return to R&D. In this approach, the rate of growth of output

or total factor productivity

is related to the R&D intensity. All the results reported by these studies (except the ones of Griliches and Mairesse (1984) and Bartelsman et al. (1996)) are based on pooled first differenced regressions, i.e. total first difference estimates. As pointed out by Mairesse and Mohnen (1995), the interpretation of this kind of estimates lies between the cross sectional and temporal dimensions of data, since the growth of total factor productivity is related to the firm’s R&D intensity whose variability is much higher across firms than over time. Let us also remind the two main differences between this specification of the productivity model and the one that estimates the elasticity of the stock of R&D capital. The former formulation assumes first, that it is the rate of return to R&D which is constant across firms and not the elasticity of the R&D capital stock, and, second, that the rate of depreciation of this capital is negligible.

Four main conclusions emerge from the estimates reported in Table 5.3. First, the results give evidence of a positive relationship between the intensity of R&D investment and the growth of total factor productivity. Indeed, the average estimated rate of return to R&D (performed on the basis of the estimates in Table 5.3 that are statistically significant) is about .29 with a lower bound of .07 (study 31) and an upper bound of .69 (study 36). Second, we can conclude from these estimates that similar downward biases as the ones observed in Tables 5.1 and 5.2 arise when no correction for the double counting of R&D is allowed for (see study 40) and when industry dummies are introduced in the productivity model (see studies 28, 34, 35, 36). A third outstanding fact emerging from Table 5.3 is the rather comparable rates of return to R&D obtained for firms of different countries. In particular, the studies 34 and 35 estimate the contribution of the R&D intensity of French, Japanese and US firms on the growth rate of sales by controlling for specific industry effects. These studies indicate that the rate of return to R&D do not differ to a great extent across countries (estimated rate of return to R&D about .31 for French firms, .30 for Japanese ones and .27 for US firms). Yet, a somewhat lower value is reported for US firms in studies 32 and 34 (estimates of .18 and .19 respectively). Finally, for some studies it is possible to compare the estimates of rates of return to R&D with the elasticities of R&D of Tables 5.1 and 5.2. As can be seen from the study of Hall and Mairesse for instance (study 40), the estimated rate of return (value of .27) is consistent with the one derived from the R&D elasticity (value of .22).

90 Sales for 5 studies and value added for 2 studies out of 17.

91 TFP-sales and TFP-value added for 5 studies respectively.

Table 5.3

Selected Productivity-R&D studies at the micro level: rates of return to R&D^a

# study data specification^b Econometric

model

est.^c s.e.^d

26 Minasian (1962) 47-57, 18 US firms, Chemicals TFP-VA F.D. Pool calc. .552

27 Mansfield (1980) 60-76, 16 US firms, Petrol. & chemicals TFP-VA F.D. Pool .275 ^(.067)*

28 Odagiri-Iwata (1986) 66-73, 135 Japanese firms TFP-VA F.D. Pool .201 ^(.109)**

" " TFP-VA-ID " .170 ^(.135)

" 74-82, 135 Japanese firms TFP-VA " .169 ^(.059)*

" " TFP-VA-ID " .113 ^(.059)**

29 Griliches-Mairesse (1984) 66-77, 133 US firms Sales Pool, calc. .35

" 66-77, 133 US firms Sales Within, calc. .64

30 Odagiri (1983) 69-81, 247 Japanese firms Other TFP-Sales F.D. Pool -.475 ^(.295)

" 69-81, 123 Japanese firms Scientific " " .256 ^(.096)*

31 Link (1981a) 71-76, 174 US firms TFP-VA F.D. Pool -.00 ^(.03)

" 71-76, 19 US firms, Transport " " .15 ^(.21)

" 71-76, 33 US firms, Chemicals " " .07 ^(.03)*

" 71-76, 34 US firms, Mechanical " " .05 ^(.07)

32 Clark-Griliches (1984) 71-80, 924 US plants Sales-ID F.D. Pool .18 ^(.05)*

" " TFP-Sales-ID " .20 ^(.05)*

33 Lichtenberg-Siegel (1991) 72-85, 5240 US firms TFP-Sales-ID F.D. Pool .132 ^(.021)*

34 Griliches-Mairesse (1983) 73-78, 343 US +185 French firms Sales F.D. Pool .28 ^(.06)*

" " Sales-ID " .12 ^(.06)*

" 73-78, 185 French firms Sales " .31 ^(.07)*

" 73-78, 343 US firms " " .19 ^(.11)**

" 73-78, 47 French firms, Drugs " " .27 ^(.15)**

" 73-78, 57 US firms, Drugs " " .41 ^(.23)**

" 73-78, 30 French firms, Chemicals " " .00 ^(.23)

" 73-78, 62 US firms, Chemicals " " -.19 ^(.36)

" 73-78, 37 French firms, Electronics " " .12 (.11)

" 73-78, 65 US firms, Electronics " " -.06 ^(.19)

" 73-78, 34 French firms, Electr. equipm. Sales F.D. Pool .45 ^(.24)**

" 73-78, 47 US firms, Electrical equipment " " -.44 ^(.33)

" 73-78, 39 French firms, Machinery " " -.55 ^(.38)

" 73-78, 112 US firms, Machinery " " .11 ^(.27)

35 Griliches-Mairesse (1990) 73-80, 406 Japanese firms Sales F.D. Pool .56 ^(.23)*

" " Sales-ID " .30 ^(.05)*

" 73-80, 525 US firms Sales " .41 ^(.09)*

" " Sales-ID " .27 ^(.10)*

" 73-80, 406 Japanese firms Sales F.D. Pool, calc. .20

" 73-80, 525 US firms Sales " .25 ^(.09)*

36 Sassenou (1988) 73-81, 394 Japanese firms Sales F.D. Pool .69 ^(.19)*

" " VA " .22 ^(.11)*

" " VA-ID " -.02 ^(.07)

37 Link (1983) 75-79, 302 US firms TFP-Sales F.D. Pool .06 ^(.04)

38 Goto-Suzuki (1989) 76-84, 13 Japanese firms, Drugs TFP-VA F.D. Pool .42 ^(.118)*

" 76-84, 5 Japanese firms, Electrical " " .22 ^(.094)*

" 76-84, 3 Japanese firms, Motor vehicles " " .33 (.138)*

39 Fecher (1990) 81-83, 292 Belgian firms TFP-Sales-ID F.D. Pool .04 (.059)

" 81-83, 113 Belgian firms, Scientific " " .05 ^(.042)

40 Hall-Mairesse (1995) 80-87, 197 French firms VA F.D. Pool .231 ^(.053)*

" " VA-DC " .273 ^(.059)*

" " VA-DC Within-calc. .22

" " VA L.D. .036 ^(.053)

" " VA -DC " .065 ^(.060)

41 Bartelsman & al. (1996) 85-89, 209 Dutch firms VA-DC L.D. .218 ^(.085)*

" 89-93, 159 Dutch firms VA-DC " .173 ^(.082)*

notes: a) adapted and extended from Mairesse and Sassenou (1991) and Mairesse and Mohnen (1995)

b) TFP = total factor productivity; VA = value-added; ID = inclusion of industry dummies; DC = correction for the double counting of R&D; F.D. = first difference; L.D. = long difference; calc. = calculated

c) R&D estimated coefficient

d) standard error, *(**) = statistically significant at the 5% (10%) level.