This article was downloaded by: [University of Saskatchewan Library]

On: 08 October 2012, At: 13:07

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Economics of Innovation and New

Technology
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/gein20

The role of research and innovation in

promoting productivity in chile
a
Jose Miguel Benavente
a
Department of Economics, University of Chile, Diagonal
Paraguay 257 of 1503, Santiago, Chile

Version of record first published: 25 Jan 2007.

To cite this article: Jose Miguel Benavente (2006): The role of research and innovation in
promoting productivity in chile, Economics of Innovation and New Technology, 15:4-5, 301-315

To link to this article: http://dx.doi.org/10.1080/10438590500512794

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-

conditions

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation
that the contents will be complete or accurate or up to date. The accuracy of any
instructions, formulae, and drug doses should be independently verified with primary
sources. The publisher shall not be liable for any loss, actions, claims, proceedings,
demand, or costs or damages whatsoever or howsoever caused arising directly or
indirectly in connection with or arising out of the use of this material.
 Econ. Innov. New Techn., 2006, Vol. 15(4/5), June/July, pp. 301–315

THE ROLE OF RESEARCH AND INNOVATION IN

PROMOTING PRODUCTIVITY IN CHILE
JOSE MIGUEL BENAVENTE∗
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

Department of Economics, University of Chile, Diagonal Paraguay 257 of 1503, Santiago, Chile

(Received December 2001; Revised July 2002; In final form November 2002)

This paper continues the empirical research line started by Crepon et al. (Crepon, 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(2), 115–158.) about the impact of research and development on innovation and innovation
on productivity of firms. In this paper, we estimate a structural model using asymptotic least squares, which corrects
for selectivity and simultaneity biases taking into consideration the particular characteristics of the available data.
We find that most of the Schumpeterian hypotheses are confirmed: research and innovative activities are related with
firm size and market power. However, in the case of Chile, firms’ productivity is not affected by innovative results,
nor by research expenditures in the short run.

Keywords: R&D; Innovation; Productivity; Asymptotic least squares; Structural estimation and censored data

JEL Classification: C13; O32

1 INTRODUCTION

This paper continues the empirical research line started by Crepon et al. (1998) about the
impact of research and development (R&D) on innovation and innovation on productivity of
firms. It is based on a model which takes into account the whole process of innovation that
includes decisions of firms to engage in R&D activities, the results of these efforts and their
impact on productivity. In this paper, we study a less developed country’s (LDC) case: Chile,
for which we have data about innovation, R&D and production activities and its results during
the period 1995–1998.
This is the first attempt to analyze research and innovation activities in an LDC and their
further impacts on productivity. As previous studies, this study presents two main characteris-
tics. On the one hand, it focuses both on innovation input (e.g. R&D) and on innovative output
like innovative sales. That is, firms invest in R&D aiming either to launch new products and/or
reduce their production costs. These, if successfully accepted by the market, could have an
impact on the firms’ production performance and economic viability. The model takes into
account the allocation of resources to R&D activities, the results of the innovation process
(where R&D expenditure is one of its determinants), and productivity, where together with
capital and labor, innovation performance is also included as one of its determinants.

∗ E-mail: [email protected]

ISSN 1043-8599 print; ISSN 1476-8364 online © 2006 Taylor & Francis
DOI: 10.1080/10438590500512794
 302 J. M. BENAVENTE

On the other hand, given the nature of the data (censored, by interval, truncated together with
some already known problems of selectivity and simultaneity), we have followed the steps
of Crepon et al. (1998) by adopting a new technique on the analysis of the data. We rely on
generalized tobit estimation to deal with problems of selectivity and on an ordered probit in the
case of interval data, all this implemented in a two-stage estimation procedure to account for
simultaneity biases in the system. The first stage consists of a method of moments estimation,
whereas the second stage uses the asymptotic least squares (ALS) to obtain consistent estimates
of the structural parameters.
The assumptions made for the estimation are rather reasonable, and we have tested some
of them using the latest available techniques. However, the main drawbacks of this study are,
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

first, the cross-sectional nature of the data and, second by, that one of the dependent variables,
R&D expenditures, is measured as a flow variable.
The paper is organized as follows. We start by presenting the basic statistics of the Chilean
case together with the main hypotheses that will be tested. The presentation of the empirical
framework, the equations and variables is the subject of Section 3. In Section 4, we present the
results of the estimation and its discussion. Concluding remarks and suggestions for further
research are the subjects of Section 5. An appendix gives details of the construction of the
data, main sources and definition of the sample, following by the comparison of our results
with those obtained by using traditional econometric techniques.

2 BASIC STATISTICS AND MAIN HYPOTHESES

We start by presenting the main characteristics of firms that we have surveyed. We report
results for the sample of firms surveyed and extrapolate these results to the whole universe of
manufacturing firms using sample weights (inflation factors).
First, we present some comments on the research and innovative variables. According to
the Chilean Innovation Survey1 , about 45% of the firms claim that they spend resources on
R&D activities. However, once the inflation factors are applied, only about 15% of the total are
engaged in R&D activities. This latter figure seems more realistic for an LDC. This significant
difference is explained by the fact that the sample was designed in such a way that firms which
represent either main actors in their markets or are considerably larger (in terms of number of
workers) compared with the rest of their competitors have an inflation factor of one. This can
be clearly observed in the average number of workers in the innovation sample and also in
terms of market share, where firms in this sample have almost three times more market power
and size than the average firm in the universe of plants (Tab. I).
When firms were asked about the introduction of a new product and/or a new process into
the market, 30% of them claim to have done this in the last 3 years. But once extrapolated to
the whole universe of firms in Chile, this figure is reduced to 18%.2
In terms of the share of innovative sales in all the sales,3 Chilean results show that they are
heavily skewed. Almost half of the firms declare that innovative sales represent less than 10%
of the total sales. Also, only 20% of the plants answered that innovative sales represent more
than 30% of their sales during 1998. On the other hand, when managers were asked about

1See Benavente & Crespi (1999) for a detailed analysis of this survey.
2A similar study of the French industry (Crepon et al. 1998) shows that on average, they have roughly the same
market share but are less diversified than their Chilean counterparts. In terms of size, the Chileans’ are slightly bigger
and, surprisingly, there exist more firms in Chile that report R&D activities (around 15%) than in France (10%).
3 As is explained in Appendix A, the average market share and the diversification index were obtained from the
firms’ sales desegregated by product lines as given by the Chilean Industry Survey Encuesta Nacional Industrial Anual
(ENIA).
 RESEARCH AND PRODUCTIVITY IN CHILE 303

TABLE I Descriptive statistics of the innovation survey and its expansion to the whole of Chilean
industry. selected indicators.

Sample adjusted by
Sample data (488 plants) weights (4084 plants)
Sample Mean Percentage Mean Percentage

R&D expenditure per employee (1998)a 292 44.7 280 15.2

Number of employees (1997) 291 100 86 100
Market share (1997, %) 0.34 100 0.13 100
Equivalent number of activities (1997) 2.15 59.2 2.17 98.5
Introduction of innovation (1998, %) 29.5 id. 17.9 id.
Value added per employee (1997)a 42,929 100 10,497 100
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

Physical capital per employee (1997) a 61,198 100 12,972 100

Engineers/employment (1997, %) 20.72 100 18.34 100
Administrative/employment (1997, %) 3.39 79.1 3.14 32.4
a
Thousands of 1998 Chilean Pesos CHP.

the importance of demand-pull and technology-push factors in the innovation dynamics of the
firms, it appears that both factors are of considerable importance in the innovation dynamics of
Chilean firms. Almost two-thirds of the managers surveyed consider that these factors affect,
from moderately to strongly, their propensity to innovate, where technology-push indicators
are much stronger than demand-pull elements. This result is to be expected in an LDC where
important channels of diffusion of technology are through FDI, purchasing of new machinery
and the introduction of new inputs usually developed overseas and where demand elements,
like customer demands, are less relevant.
Given the available data, we aim to test most of the traditional hypotheses related with
industrial innovation.4 These studies stressed the favorable position that for innovation could
have for large firms with a good degree of market share. This by exploiting not only their size
but also their market power.
On the one hand, market concentration allows the existence of monopolistic rents that
enable firms within the industry to finance R&D projects. In a perfectly competitive world,
private investment in R&D would not be feasible. Additionally, when imperfections in capital
markets are recognized, large firms tend to have greater capabilities of securing the necessary
resources to finance R&D projects, barring uninsurable risk. Moreover, the potential total
impact of the results of an R&D project can be significantly greater in large firms, given larger
sale volumes. Hence, a product or process innovation that allows for an increase in the price-
cost margin will have greater absolute effects on profits on a firm with greater sales. Finally,
larger plants have greater absolute incentives to improve internal process technologies, which
in conjunction with economies of scale should lead firms to make greater relative efforts on
process innovations than their smaller counterparts.
However, there exist several external forces which stimulate firms to innovate despite
their size and market power. The first of such forces is based on demand factors, such as
market growth. Schmookler (1966) first formulated such rationale known as the ‘demand-pull
hypothesis’. However, the role of scientific advancements in stimulating industrial innovative
efforts may influence the path and rate of technology advance. The rationale was that advances
in science enabled ‘technology-push’-based innovations5 through the development of new
concepts or when incorporated in new machinery and/or inputs. The relative opportunities
to innovate within a given industry on the basis of scientific progress form the basis for the
notion of technology opportunities.

4 For a survey of them see Cohen & Levin (1989), Cohen (1995), and Cohen and Klepper (1996).
5 First mentioned by Rosenberg (1974).
 304 J. M. BENAVENTE

The empirical estimation procedure aims to capture the effect of some of these variables over
the research and innovation effort of firms. It consists of four main equations: two for R&D,
one for innovation and one for productivity. Given the design of the Chilean Innovation Survey,
some restrictions are imposed on the data which require a different econometric treatment. We
will estimate a reduced version of the model which includes only firm’s characteristics such as
size and market share, and an expanded version including demand-pull and technological-push
variables.

3 EMPIRICAL MODEL AND STRATEGY FOR ESTIMATION

Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

In this section, we empirically model the whole innovation process. We assume that there exists
a set of firm’s characteristics, market structure and technological variables, which conditions
and shapes investments in R&D as well as the successful introduction of new products and
processes into markets.
As in Crepon et al. (1998), we model the process of innovation with four main equations:
two for R&D, one for innovation and one for productivity. It is important to note that each
equation requires a different econometric treatment depending on the characteristics of the
data. To make this work comparable to the French case, we will estimate a reduced version
of the model which includes only firm’s characteristics such as size and market share, and
an expanded version including demand-pull and technological-push variables. The definition
of variables and the econometric specification of each equation are the subjects of the next
sections.

3.1 Research Equations

We will rely on a generalized tobit framework to model the decision to invest and the amount
invested in research activities. It consists of two separate equations: one related to the decision
to invest in R&D and the other to the amount of resources involved.
More precisely, we assume that there exists a latent dependent variable gi∗ for the firm i
given by the following equation:
gi∗ = xi0 b0 + ui0 (1)

where xi0 is a vector of explanatory variables, b0 the associated coefficient vector and ui0
an error term, and gi∗ represent some decision criterion like the expected present benefits
associated with a research project. The econometrician observes that resources are invested in
R&D activities if gi∗ is positive or bigger than a given threshold.
We then assume that a latent or true intensity of research ki∗ for firm i is determined by a
second equation:
ki∗ = xi1 b1 + ui1 (2)

where ki∗ = ki is the current R&D expenditure of firm i when it is engaged in research activities,
i.e. when gi∗ is bigger than the minimum industrial threshold. Here, ki∗ and ki are both expressed
in logarithms, xi1 is a vector of the explanatory variables, b1 the associated coefficient vector
and ui1 an error term that summarizes omitted determinants and other sources of unobserved
heterogeneity.
Given that ki∗ is only observable when gi∗ is bigger than the minimum threshold, we need to
specify their joint distribution in order to obtain an estimable model. Following the standard
literature about generalized tobit models, we assume that the error terms in Eqs. (1) and (2)
 RESEARCH AND PRODUCTIVITY IN CHILE 305

follow a joint normal distribution:




2 
ui0 iid 0 σ0 ρσ0 σ1
−→ N , (3)
ui1 0 ρσ0 σ1 σ12

where σ0 and σ1 are the standard errors of ui0 , and ui1 , respectively, and ρ is their correlation
coefficient.
In the implementation of the model, we use a flow variable for research expressed in thousand
Chilean pesos per worker, given our available data. We are particularly interested to test the
impact of firm size and market power on the probability as well as on the amount of resources
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

invested in research activities. More precisely,

xi0 = xi1 = (li , siw , di , δi1 , δi2 , δi3 , τi1 , τi2 , τi3 , Si1 , . . . , Si9 )

where li is employment, siw the 1-year lagged average market share and di the equivalent
number of industry segments, all the three variables expressed in logarithms. As in previous
works, δk and τk with k = 1, 2, 3 are two sets of demand-pull and technological-push dummies.
j
Finally, the Si are nine industry dummies equal to 1 if firm i belongs to industry j and 0
otherwise. These later dummies replace the constant term so that each industry has a different
intercept. This is consistent with the design of the survey, which contains strata for different
productive sector up to two digits ISIC (i.e. nine sectors).
As suggested in the last section, we expect a positive impact of size on investments in
R&D once controlled for industry’s characteristics. On the other hand, a positive relationship
between market power and the probability of reporting R&D expenditure is expected. We also
expect an impact from technological-push and demand-pull variables on research investments
once controlled for other characteristics of the industries. In this later case, we are particularly
interested in the significance of the impact rather than the sign because given the design of
the variables negative impact is unfeasible.

3.2 The Innovation Equations

In general, it is assumed that innovative output is related to enhancements in the productivity
capacity of the firm. However, in empirical work, there exist several ways to proxy innovation
output. More common proxies are the number of patents or share of innovative sales. The
Chilean Innovation Survey only provides information on the latter because patents are very
scarce. The firm is asked what is the percentage share of its 1998 sales coming from products
launched in the last three years 1996–1998 and/or what are the results of the implementation
of absolutely new production techniques during the same period. The firm has to answer on a
four-point scale: 0–10, 10–30, 30–70 and 70–100%.
We will assume that the underlying (unobserved) true share of sales ti∗ can be viewed as an
innovation intensity variable where innovation is measured by the number of sales-weighted
innovations. Because the share is known only by intervals, we can specify the innovative sales
equation as an ordered probit model:

ti∗ = αk ki∗ + xi2 b2 + ui2 (4)

where ti∗ is the underlying (unobservable) true share expressed as a logarithm, xi2 is a vector
of explanatory variables, b2 the associated coefficient vector and ui2 an error term which we
will assume to be normally distributed with mean zero and variance σ22 . The coefficient αk
is the elasticity of the percentage share of innovative sales relative to research expenditure,
 306 J. M. BENAVENTE

a measure of the return or impact of research on innovation output, and x2 is a vector of

exogenous variables:

xi2 = (li , δi1 , δi2 , δi3 , τi1 , τi2 , τi3 , Si1 , . . . , Si9 )

with the same notation as above. Note that it is assumed that market share and diversification
do not enter directly in the innovation equation but only indirectly through research in order to
identify the system. By contrast, we assume that demand-pull and technological-push factors
could affect innovation output directly and indirectly. Also, including size in the innovation
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

equation allows us to test whether the effect of firm size on innovation passes completely
through the size of research activities.
As in the research equations, we include sectoral dummies which capture technological
differences between markets not captured by technology-push or demand-pull variables.

3.3 The Productivity Equation

The last equation of the system is for labor productivity. Following the standard litera-
ture, we assume a Cobb–Douglas production function with physical capital and employment
augmented with skill composition and innovation output. Then we have:

qi = αI ti∗ + xi3 b3 + ui3 (5)

where qi stands for labor productivity measured as a logarithm. The vector of the factors other
than innovation output is:

xi3 = (li , ci , Ei , Ai , Si1 , . . . , Si9 )

with ci being the physical capital per employee as a logarithm and Ei and Ai being the share of
engineers and administrators, respectively, in the total number of employees. The coefficient
αI is the elasticity of total factor productivity with respect to innovation output and b3 is
a vector of coefficients related to the elasticity of scale, that of physical capital and the skill
composition parameters, reflecting percentage differences in efficiency of skilled labor relative
to unskilled labor, respectively.6

3.4 The Overall Model and its Estimation

All four equations form a recursive non-linear system where innovation output is proxied by
the ratio of innovative sales over total sales. The first two equations concern the R&D decision

6 Following Crepon et al. (1998), the interpretation of the parameters of skill composition is the following.
Assume that labor, corrected for quality, enters the production function as L∗ = fU LU + fE LE + fA LA instead
of the uncorrected total employment L = LU + LE + LA , where U stands for unskilled, E for engineers and
A for administrators. Then the first expression can be re-written as L∗ = fU (L − LE − LA ) + fE LE + fA LA
or as L∗ = fU L{1 + (fE /fU − 1)LE /L + (fA /fU − 1)LA /L}. Using logarithm on both sides results in ln L∗ ∼ =
ln fU + ln L + uE E + uA A and thus the coefficients of E and A in the production function are, respectively,
b3E ≈ (∂q/∂l)uE and b3A ≈ (∂q/∂l)uA .
 RESEARCH AND PRODUCTIVITY IN CHILE 307

and expenditure:

gi∗ = xi0 b0 + ui0 (6)

ki∗ = xi1 b1 + ui1 (7)

whereas the two last equations, for innovation output and productivity, are the following:

ti∗ = αI ki∗ + xi2 b2 + ui2 (8)

qi = αk ti∗ + xi3 b3 + ui3 (9)
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

No restrictions are imposed on the correlations between the disturbances ui0 , ui1 , ui2 and ui3 .
However, we want to take into account the problems of selectivity and simultaneity that can
arise because of the nature of the data (truncated, censored, discrete and interval) and the
endogeneity structure of some of the equations (research expenditure is endogenous in the
innovation equation and innovation is also endogenous in the productivity equation).
There are several ways to estimate this system of equations. As suggested by Crepon
et al. (1998), maximum likelihood (ML) would be impractical due to the non-closed form of
the joint distribution. On the basis of the studies of Lee (1981), they favor the use of the ALS
estimator, also known as the minimum distance (MD) estimator (Gouriroux & Monfort, 1996).
This estimator is relatively more efficient and its computational costs are lower compared with
alternative estimators like the GMM.

4 RESULTS

4.1 Results of the Basic Model

Table II presents the results using ALS for the basic model. Results obtained with more
traditional econometric techniques like OLS, 2SLS, ML and ALS for the equations estimated
separately are presented and discussed in Appendix B.
The first two columns of Table II show the estimates of the research equations and the last
two the estimates of the innovative sales equation and the productivity equation with innovative
sales. The bottom line shows the results of Bera et al. (1984) for normality in the probit model
and that of the Pagan and Vella (1989) test for normality in the tobit model.

TABLE II Results for basic model.

Non-zero Log research Log Log value

research expenditures per innovation added per
Dependent variable expenditures employee (ki ) intensity (ti ) employee (qi )

Estimation method Probit Tobit Intensity Productivity

R&D expenditure per employee ki – – −0.076 (0.199) –
Innovation intensity ti – – – 0.170 (2.072)
Number of employees li 0.182 (0.055) 0.082 (0.129) 0.142 (0.029) 0.100 (0.185)
Market share si 0.139 (0.049) 0.195 (0.121) – –
Equivalent number of activities di −0.025 (0.130) −0.078 (0.276) – –
Physical capital per employee ci – – – 0.686 (0.044)
Statistical Test (Pagan–Vella,
Bera–Jarque–Vella) 1.57 (0.45) 0.103 (0.92)

Note: Optimal ALS with eight industry dummies. Asymptotic standard errors in parentheses. Sample of 438 plants.
 308 J. M. BENAVENTE

Starting with the decision to allocate resources in R&D activities (probit column), firm
size has a positive and significant impact on this probability. This increase with size is a well-
documented fact in the empirical literature. After controlling for size and sector, it appears that
the probability of doing R&D also increases with the degree of market share. As suggested by
Crepon et al. (1998), this is a new link that they also found in their analysis of French firms.
The analysis with R&D intensity equation shows that firm’s size has no significant impact
on this variable, suggesting that the elasticity of R&D expenditure to size is one. This is
consistent with previous work in this line of research.7 However, contrary to the French case,
in our study with Chilean firms, neither market share nor firm’s diversification has a significant
impact on R&D intensity. It seems that these two variables only affect the decision to enroll
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

in R&D activities but not the size of it. This last fact is also consistent with previous work
summarized in Cohen and Levin (1989) and Cohen (1995).
Regarding the innovation equation, current R&D expenditures have no significant impact
on current sales-weighted innovations once controlled by size and sector. This result can be
explained by the dynamic nature of the innovation process where it is difficult to expect instant
success related to R&D activities. However, in the case of Chile, it appears that bigger firms
tend to have more success in their innovative sales compared with their smaller counterparts.
Finally, estimates for the productivity equation shows that there exist constant returns to
scale and a physical capital elasticity of about 0.7. However, current sales-weighted innovations
which proxy innovation efforts have no statistical significance in the explanation of producti-
vity of firms.8 This result may be due to the flow characteristic of the Chilean innovation data.

4.2 Results of the Extended Model

The ALS estimates of the extended version of the model are presented in Table III. We have
included six additional rows in the research and innovation equations (for the weak, moderate
and strong type of answers) which define our demand-pull and technology-push indicators,
and two additional rows in the productivity equation for our two skill composition variables:
the share of engineers and administrators relative to other categories of employees.
With respect to the variables already discussed, they do not change dramatically between
either specification. The only exception is firm size, which appears to be significant in the
productivity equation but not in the innovation one.
Both market demand and technological opportunities, measured here by demand-pull
and technology-push indicators, appear to have positive effects on the firm’s probability of
engaging in R&D activities and their intensity. However, only the proxies for technological
opportunities have a significant effect in the Chilean plants. In the case of research intensity,
these effects tend to increase across the type of answers as in the French case. Surprisingly,
in the case of the probability equation, technology-push variables are inversely related with
probability. That is, the likelihood that firms allocate resources to R&D activities does not
increase if managers perceive that technological opportunities are important; moreover, it may
decrease. These estimators are, however, positive and statistically significant. In the case of the
innovation equation, neither demand-pull nor technology-push indicators have a significant
effect on the sales-weighted innovations.
Finally, our two skill composition variables are significant and with large positive coeffi-
cients in the productivity equation.9 Especially, administrative workers, and in a second

7 Known as Stylised Facts 2 and 3 under Cohen and Kepler’s (1996) terminology.
8 As can be observed in Table B.3 in Appendix B, a similar result is obtained when R&D expenditure is used
instead.
9 Also they are not correlated with knowledge capital proxies when comparing both specifications, suggesting
neither complementarity nor substitution between these factors.
 RESEARCH AND PRODUCTIVITY IN CHILE 309

TABLE III Results for the extended model.

Non-zero Log research Log Log value

research expenditures per innovation added per
Dependent variable expenditures employee (ki ) intensity (ti ) employee (qi )

Estimation method Probit Tobit Intensity Productivity

R&D expenditure per employee ki – – −0.181 (0.344) –
Innovation intensity ti – – – 0.179 (0.113)
Number of employees li 0.077 (0.059) −0.115 (0.130) −0.011 (0.064) 0.140 (0.041)
Market share si 0.126 (0.052) 0.143 (0.116) – –
Equivalent number of activities di −0.097 (0.132) −0.153 (0.262) – –
Physical capital per employee ci – – – 0.670 (0.021)
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

Engineers/personnel Ei – – – 0.652 (0.177)

Administrative/personnel Ai – – – 5.952 (1.459)
Demand-pull:
Weak (δi1 ) 0.191 (0.273) 0.035 (0.542) 0.360 (0.280) –
Moderate (δi2 ) 0.286 (0.234) 0.157 (0.451) 0.085 (0.247) –
Strong (δi3 ) 0.292 (0.220) 0.111 (0.430) 0.335 (0.227) –
Technology-push:
Weak (τi1 ) 0.980 (0.312) 1.778 (0.641) 0.890 (0.671) –
Moderate (τi2 ) 0.904 (0.255) 2.011 (0.540) 1.139 (0.735) –
Strong (τi3 ) 0.895 (0.244) 2.223 (0.522) 1.402 (0.795) –

Note: Optimal ALS with eight industry dummies. Sample of 438 plants.

place, engineers, make a massive difference in their impact on the plant’s productivity
compared with other type of workers. This is obviously reflected in their relative salaries.10

5 CONCLUSIONS

One of the main characteristics of empirical studies on R&D and innovation is the particular
structure of the empirical models adopted. Most of the research in this area has been conducted
by using a simple equation relating a measure of research effort to firm’s characteristics and/or
market structure.
There exist several biases that arise by using single-equation models. Probably, the most
important is the selectivity one, which is the result of using samples, including firms with
no reported R&D expenditure. Recent work has tackled this problem by estimating a system
of equations in which one equation deals with the selectivity issue and the other with the
research intensity. A generalized tobit procedure is the adequate tool to estimate these systems
in a consistent and efficient manner. However, we are dealing with only part of the whole
innovative process that takes place at the level of the firm. One of the main objectives of this
paper is to explore the links between R&D and innovation, and the impact of them on firms’
productivity for an LDC case.
Results show that when all the steps – R&D, introduction of new products and processes,
and firm’s productivity – in the innovative process are incorporated, most of the traditional
hypotheses of firm’s characteristics and market structure are validated under this broader
set-up.
Indeed, in the case of Chile, firm size is related to the probability that firms are engaged
in research activities. However, size is not related to the amount of resources allocated for
these activities once controlled for sectoral differences, suggesting a constant return to scale

10 See INE (1998).

 310 J. M. BENAVENTE

in research investments. Results also show that technological opportunities do play a major
role in research activities especially when innovative ideas are embedded in new machinery
and output.
There is a constant return to scale in the productivity equation, and after including labor
skills in the estimation of the productivity, both engineers and administrative shares have a
positive and significant effect.
The econometric methods used take into account not only the particular characteristics of
the data such as truncation, interval and censored but also the selectivity and simultaneity
problems that are embodied in this kind of empirical exercise.
However, some of the findings of our study of the Chilean case were unexpected. Among
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

them the fact that neither research expenditure nor innovation has a significant impact on
innovation sales and productivity, respectively. These results can be explained partly by the
implicit assumption in the model that there are no lags between the implementation of innova-
tions and impact on productivity, a subject widely discussed in the literature. However, a more
plausible explanation is related with the fact that productivity is measured as value added per
worker. If innovation is mainly related with embodied technical change, then this effect is not
captured in this model. Results in related work11 suggest that machinery is related with the
probability to report R&D expenditures, but we do not have the data to test this hypothesis for
productivity.
We intent to tackle this problem in the near future when more information for the Chilean
firms will be available.

Acknowledgements
This paper is based on Chapter 6 of my DPhil thesis at Oxford. I wish to thank Bronwyn Hall,
Paul David, Nick von Tunzelmann, Jacques Mairesse, Emmanuel Duguet, Valpy FitzGerald,
participants at the Gorman workshop at Oxford, and two anonymous referees for their
comments and suggestions.

References
Benavente, J.M. (2002) Determinants of Industrial Research and Innovation: The Case of Chile. Unpublished DPhil
Thesis, University of Oxford.
Benavente, J.M. and Crespi, G. (1999) Informe Final: Encuesta sobre la Innovacion Tecnologica en la Industria
Manufacturera Chilena. Manuscript, University of Chile.
Bera, A., Jarque, C. and Lee, L.F. (1984) Testing the Normality Assumption in Limited Dependent Variable Models.
International Economic Review, 25, 563–578.
Cohen, W. (1995) Empirical Studies of Innovative Activity. In Stoneman, P. (ed.) Handbook of the Economics of
Innovation and Technological Change. Oxford: Blackwell.
Cohen, W. and Levin, R. (1989) Empirical Studies of Innovation and Market Structure. In Schmalensee, R. and Willig,
R. (eds.) Handbook of Industrial Organization. Amsterdam: North Holland.
Cohen, W. and Klepper, S. (1996) A Reprise of Size and R&D. The Economic Journal, 106, 925–951.
Crepon, 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(2), 115–158.
Gourieroux, C. and Monfort, A. (1996) Statistics and Econometric Models. Cambridge: Cambridge University Press.
INE (1998) Ingreso de Hogares y Personas. Encuesta Suplementaria de Ingresos. Instituto Nacional de Estadisticas,
Santiago de Chile.
Lee, L.F. (1981) Simultaneous Equations Models with Discrete and Censored Variables. In Manski, C.F. and
McFadden, D. (eds.) Structural Analysis of Discrete Data with Econometric Applications. Cambridge, MA:
MIT Press.
Pagan, A. and Vella, F. (1989) Diagnostic Test for Model Based on Individual Data: A Survey. Journal of Applied
Econometrics, 4, S29–S59.
Rosenberg, N. (1974) Science, Innovation and Economic Growth. Economic Journal, 84, 90–108.
Schmookler, J. (1966) Invention and Economic Growth. Cambridge, MA: Harvard University Press.

11 Benavente (2002).
 RESEARCH AND PRODUCTIVITY IN CHILE 311

APPENDIX A

A.1 The Survey

The data used in the first part of this study comes from the first Technological Innovation in the
Chilean Manufacture Survey held during 1998. It was a joint work between the National Insti-
tute of Statistics (INE) and the Ministry of Economics. The survey was conducted according
to the OECD instructions defined in the Oslo and Frascati Manuals.
In general, the survey was designed to capture qualitative variations in the innovation activ-
ities of firms. The main objective was to identify the factors underlying these changes and to
evaluate the efficiency of public policy. Therefore, the survey contains only a limited number
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

of measures about the levels of innovation activity within firms, with the only exception of
expenditures on R&D and other innovation activities like patents, royalties and training in
R&D. However, because the sample of firms was obtained from the universe of firms con-
tained in the Annual Industrial Survey, other quantitative characteristics of the firms can be
included and analyzed (e.g. employment, exports, investment).
A second feature of the design of the survey is the procedure to estimate variations in the
innovation activity. A set of questions was asked to capture perceptions from engineers and
production managers of the plants. Therefore, the measurement of the changes in the innovation
dynamics was made by those directly involved in the production units. To standardize the
answers, they were ordered according to the intensity or importance given to each question
in those cases where an innovation was present. The rank varies from 1 to 4 (from lower to
higher importance).
A third aspect to highlight is that although the survey contains only a single observation in
time, some dynamic behavior can be obtained from it. The questions were designed to capture
information on productive performance during the last 3 years rather than the year in which
the survey was conducted.
There are some limitations to be taken into account. First, there is room for arbitrariness in
the intensity scaling and its interpretation. Secondly, the problem of aggregating answers that
are qualitative in nature. In fact, from an answer of 4 (which represents that the phenomenon
was very intense or very important in the firm), it cannot be automatically inferred that the
phenomena have taken the same characteristics in all the firms that gave 4 as an answer for
the same question. An answer of 4 in two firms to the same question has to be interpreted that
the phenomenon has been perceived as important or intensive in each firm despite differences
between each case.
Finally, the evaluation of variations rather than levels of innovative activity makes the
international comparisons very difficult.

A.2 The Sample of Plants

The information referred to plants rather than firms because the survey aimed to detect inno-
vative phenomena developed at a plant level. Nevertheless, in the cases of firms with more
than one plant, the survey was applied to all the plants of the same firm in order to capture
the potential externalities and links between the plant of the same firm. A probabilistic and
stratified sampling procedure, considering the size by value added, was performed for the
study.
The design considered two groups of plants. The first is named forced inclusion strata.
The inclusion criterion here is plants with more than 2% of the value added or exports or
investment of this productive sector (two-digit ISIC). The second group is randomly selected.
The selection was made in each size strata, with a probability proportional to the value added
of each plant. The final sample included 488 plants where the method chosen to select the
 312 J. M. BENAVENTE

sample guarantees a good representation of the universe, with a sampling error of 5.1% at
95% of confidence.

A.3 The Data

Traditional Variables
The information on the firm’s current accounts and balance sheets and on the number of
employees comes primarily from the ENIA surveys (Encuesta Nacional Industrial Anual).
From them, we have constructed the firm’s value added, its fixed assets’gross book value and its
total number of employees (average over the year), and we have computed labor, productivity
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

and physical capital intensities as li = number of employees, qi = value added per employee
and ci = physical capital per employee. These variables are expressed in logarithm in our
estimations.
The information about the distribution of employees comes from the same source. The
data allow differentiation between administrators and engineers at a managerial level from
blue-collar workers at the floor level and the rest of the employees.
Average firm market share and diversification indexes are computed from the same ENIA
survey. This survey gives detailed information about the decomposition of firm’s sales in all
its different lines of business up to a Chilean equivalent of eight-digit ISIC. In our estimations,
we have defined Si,k as the sales of firm i for its product k in the industry segment or market k
 
Si = Si,k and Sk = Si,k
k i

are, respectively, the overall sales of firm i (overall its products) and overall sales on market
k (overall firms) without including its exports. The market share si,k of firm i on market k and
the share of product k on firm i total sales are thus equal to
Si,k Si,k
si,k = and bi,k =
Sk Si
Then for each diversified firm i, we can define the weighted average market share siw and the
diversification index di as
 1 
siw = bi,k × si,k and = hi = 2
bi,k
k
di k

with di being the inverse of the Herfindahl concentration index hi of the firm’s sales. For
example, a non-diversified firm, i.e. with only one k, we have siw = si and di = hi = 1. The
diversification index di for the firm i can be interpreted as the equivalent number of product
lines with equal sales. The higher this index, the higher the diversification of the firm. In the
model, we introduce variables siw and di in logarithms.
Finally, we include in all equations of the model a full set of eight industry dummies:
Si1 , Si2 , . . . , Si8 . These are defined at a higher level of classification (two-digit ISIC) than the
average market share and diversification variables on the basis of the firm’s main industrial
activity.

Research and Innovation Data

Information about R&D expenditures is taken from the ‘Encuesta de Innovacion Tecnologica
en la Industria Manufacturera Chilena’. Plants were requested to inform about their R&D
 RESEARCH AND PRODUCTIVITY IN CHILE 313

expenditures during 1998 and 1997 in 1998 Chilean pesos using the Frascati methods. This
survey has two versions, the first in 1995 and the second in 1998. Given the characteristics
of the sample design, only few firms (less than 200) were interviewed in both occasions, so
larger time series of research and innovative activities were difficult to obtain. In our model,
gi is a research dummy equal to 1 if firm i has reported R&D expenditures during 1997 and
1998 and 0 otherwise. On the other hand, ki is the log of the R&D expenditures of 1998.
The share of innovative sales is also obtained from the same survey. This variable is based
on a question asking innovating firms what percentage of their 1998 sales is imputable to
new products and/or new processes launched between 1996 and 1998. Firms answered on an
interval scale: 0–10, 10–30, 30–70 and more than 70%.
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

Firm’s demand-pull and technology-push indicators are, respectively, based on the average
answer to the following questions: do you think that the ideas for innovation come from either
research activities performed together with clients and suppliers, other firms or simply copy?,
and do you think that innovations come from concepts and ideas embodied in your purchases
of new machinery, the use of new inputs and/or the consultation of books, periodicals or
visiting expositions? Questions are ranked in a scale between 1 and 4, giving a qualification of
weak for those answers equal to 1, moderate for 2 and 3, and strong for only those answered
with a value of 4.

APPENDIX B

This appendix contains an assessment of the biases likely to arise in innovation and pro-
ductivity studies. One of the main objectives of this study is to correct for selectivity and
simultaneity biases, commonly found in applied innovation studies. We have also taken into
account the main characteristics of the available data like censoring and truncation. We also
intend to assess the magnitude of these biases while using more standard econometric meth-
ods. In this section, we present results of the estimation of the same group of equations
using ordinary least squares (OLS), two stage least squares (2SLS), ML and first-step and
second-step ALS estimators. We only perform this exercise for the basic specification of our
model to the sample of innovative plants and when appropriate for the subsample of R&D
performing firms.

B.1 Research Equations

Table B.1 compares the OLS estimates for the research intensity equation for the subsample
restricted to the R&D performing firms to the corresponding ML estimates for the generalized
Tobit specification. As can be seen, they differ substantially. Moreover, if OLS estimation was

TABLE B.1 Research and development equation dependent variable = log of R&D
spending per employee (ki ).

Variable li si di

Generalized tobit estimationa

Probit equation 0.182 (0.055) 0.139 (0.049) −0.025 (0.130)
Tobit equation 0.082 (0.129) 0.195 (0.121) −0.078 (0.276)
OLS estimation 0.297 (0.105) 0.183 (0.079) −0.030 (0.234)

Note: Number of observations is 438. All regressions include eight industry dummies. Standard
errors in parentheses.
a
Correlation among the residuals = 0.830 (0.057).
 314 J. M. BENAVENTE

TABLE B.2 Innovation intensity equation dependent variable = log of innovation

percentage in sales (ti ).

Variable ki li gi

All observations (N = 438)

OLSa −0.030 (0.046) 0.177 (0.050) 0.360 (0.223)
Two-stage least squaresb −2.766 (5.142) −0.077 (0.460) 12.830 (23.432)
Maximum likelihoodc −0.071 (0.051) 0.073 (0.056) 0.594 (0.247)
ALS (first-step: OLS) 0.105 (4.750) 0.060 (1.073) ∗
ALS (second-step: GLS) 0.133 (0.241) 0.060 (0.061) ∗

Positive R&D observations (N = 197)

OLSa 0.036 (0.056) 0.190 (0.086) ∗
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

Two-stage least squaresb 2.539 (5.046) 0.134 (0.432) ∗

Maximum likelihoodc −0.093 (0.057) −0.007 (0.089) ∗

Note: All regressions include eight industry dummies. Standard errors in parentheses.
a
The logarithm of the interval centre was taken as the dependent variable.
b
Instruments for 2SLS: li , si , di and the eight industry dummies.
c
Ordered probit with known thresholds.

used instead, we would have concluded that the market share is an important explanation in
the R&D intensity of firms and that the elasticity of R&D to firm size is bigger than one, which
both are not the case.

B.2 Innovation Sales Equation

Before discussing these results, it is important to notice that for OLS and 2SLS estimation we
have assumed that the value of innovative sales was the interval centre, that is 5, 20, 50 and
85%, respectively.
As can be observed from Table B.2 there are problems with selectivity given the differences
in the value for the estimates between the two samples. However, it has to be noted that,
particularly in the case of 2SLS, estimates are very imprecise. In contrast, simultaneity
problems can exist while comparing results from ML estimation with ALS. For example,
note that in the case of firm size, this coefficient is not significantly different from zero under
ML but is so under ALS. It seems that ML is not taking into account the effect that firm size
has through the probability to perform R&D activities into innovation sales.

TABLE B.3 Productivity regression with R&D expenditures dependent variable = log of value added per
employee (qi ).

Variable ki ci li gi

All observations (N = 438)

OLS 0.058 (0.037) 0.693 (0.050) 0.129 (0.048) −0.159 (0.184)
Two-stage least squaresa −2.603 (4.836) 1.057 (0.670) −0.325 (0.812) 11.943 (21.986)
Positive R&D observations (N = 197)
OLS 0.131 (0.040) 0.686 (0.068) 0.197 (0.078) ∗
Two-stage least squaresa 8.195 (318.88) −4.057 (187.58) 2.372 (85.932) ∗

Note: All regressions include eight industry dummies. Standard errors in parentheses.
a
Instruments for 2SLS: li , ci , si di and the eight industry dummies.
 RESEARCH AND PRODUCTIVITY IN CHILE 315

TABLE B.4 Productivity regression with innovation intensity dependent variable = log
of value added per employee (qi ).

Variable ti ci li

All observations (N = 438)

OLSa 0.060 (0.048) 0.696 (0.048) 0.121 (0.043)
Two stage least squaresa,b 0.833 (0.821) 0.619 (0.083) 0.009 (0.126)
ALS (first-step: OLS) 0.362 (35.932) 0.705 (1.000) 0.104 (2.663)
ALS (second-step: GLS) 0.032 (2.099) 0.689 (0.044) 0.109 (0.188)

Note: All regressions include eight industry dummies. Standard errors in parentheses.
a
The logarithm of the interval centre was taken as the right-hand variable.
Downloaded by [University of Saskatchewan Library] at 13:07 08 October 2012

b
Instruments for 2SLS: li , ci si , di and the 8 industry dummies.

B.3 Productivity Equation

The first interesting observation of the results presented in Tables B.3 and B.4 is that capital is
always significant and with an elasticity about 0.7 in all the equations. The exception is again
the 2SLS, which, as in the French case, estimates using this technique are quite imprecise.
Table B.3 also shows that there exist selectivity problems when only the subsample of R&D
performing firms is considered. Estimators differ substantially between both samples.
In none of the estimations, R&D expenditures and/or innovation sales has a significant
effect on the firm’s productivity. This confirms the results obtained with ALS, which corrects
for potential problems due to selectivity and simultaneity, where knowledge-related variables
are not important in explaining productivity changes, at least contemporarily, in the Chilean
manufacturing industry.