Procyclicality of R&D investments : a critical study

(1)

Faculty of Economics and Business

Amsterdam School of Economics

Requirements thesis MSc in Econometrics.

1. The thesis should have the nature of a scientic paper. Consequently the thesis is divided up into a number of sections and contains references. An outline can be something like (this is an example for an empirical thesis, for a theoretical thesis have a look at a relevant paper from the literature):

(a) Front page (requirements see below)

(b) Statement of originality (compulsary, separate page) (c) Introduction (d) Theoretical background (e) Model (f) Data (g) Empirical Analysis (h) Conclusions

(i) References (compulsary)

If preferred you can change the number and order of the sections (but the order you use should be logical) and the heading of the sections. You have a free choice how to list your references but be consistent. References in the text should contain the names of the authors and the year of publication. E.g. Heckman and McFadden (2013). In the case of three or more authors: list all names and year of publication in case of the rst reference and use the rst name and et al and year of publication for the other references. Provide page numbers.

2. As a guideline, the thesis usually contains 25-40 pages using a normal page format. All that actually matters is that your supervisor agrees with your thesis.

3. The front page should contain:

(a) The logo of the UvA, a reference to the Amsterdam School of Economics and the Faculty as in the heading of this document. This combination is provided on Blackboard (in MSc Econometrics Theses & Presentations).

(b) The title of the thesis

(c) Your name and student number (d) Date of submission nal version

(e) MSc in Econometrics

(f) Your track of the MSc in Econometrics 1

Procyclicality of R&D Investments

a Critical Study

Gwendolyn Stuitje

Student number: 6184472

Master’s programme: Econometrics

Specialisation: Free Track

Date of final version: August 15, 2017

Supervisor: dr. J. C. M. van Ophem

Second reader: mw. dr. E. Aristodemou

Abstract

While economic theory predicts that firms time R&D investments in recessionary pe-riods due to lower opportunity costs, recent empirical studies demonstrate that R&D is procyclical. This paper provides a critical analysis of one of the empirical studies into the procyclicality of R&D investment, that of Fabrizio and Tsolmon (2014). We argue that they incorrectly model R&D investment and provide an alternative model. A robustness check verifies the procyclical pattern of R&D, but casts doubt on the main prediction that obsolescence affects the timing of R&D. To obtain more insight in the timing of R&D and the sources to which its procyclical pattern is attributed, we test the hypotheses of Fabrizio and Tsolmon on different sub-samples of firms and find different results. Furthermore, this paper argues that Fabrizio and Tsolmon use an incorrect es-timation sample and therefore alters the sample. A Heckman selection model indicates the presence of a selection bias.

(2)

This document is written by Gwendolyn Stuitje who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used

in creating it. The Faculty of Economics and Business is responsible solely for the

(3)

Introduction

There are two views on the timing of R&D investments. On the one hand, prevailing economic theory predicts that firms time their R&D investments during economic reces-sions. This view relies on the idea that the opportunity costs of investing in R&D are lower during recessions due to the decrease in demand (Schumpeter, 1938). Therefore it is argued that R&D investments have a countercyclical occurrence. However, over the recent years there has been done more empirical research into nature of R&D investments and the findings of these studies have led to the opposite view that R&D investments are procyclical.

There are several reasons why R&D expenditures do not exhibit the countercyclical pattern as expected by the economic literature. In the first place, firms face financial constraints during recessions that disable them to invest in R&D or maintain their R&D investments (Aghion et al. 2010). Secondly, firms purposefully delay R&D investments until times of economic prosperity in order to maximize profits (Barlevy, 2007). This latter argument is derived from the assumption that R&D investments coincide with the introduction of new innovations, which makes new inventions susceptible to imitation as soon as firms invests in R&D. As imitation eats away profits, firms are inclined to time R&D investments during periods of high demand in order to capture the highest possible returns on their innovations. However, this argument does not hold if R&D investment and the introduction of an innovation do not happen simultaneously.

Fabrizio and Tsolmon (2014) test the latter prediction by an empirical examination of the timing of R&D expenditures of firms that which are part of industries that are more prone to imitation. As susceptibility to imitation depends on the protection of intellectual property within an industry, Fabrizio and Tsolmon create a measure for the effectiveness of patent protection within industries to carry out the abovementioned examination. Furthermore, Fabrizio and Tsolmon build on the theoretical arguments that are given to explain the procyclical pattern of R&D and provide an additional hypothesis. Using the idea that firms want to maximize profits, they argue that faster rates of product obsolescence causes firms to time R&D investments in economic booms, because just as imitation, product obsolescence reduces the rents a firm can capture from its innovation. An advantage of this hypothesis is that it does not need to make assumptions about the coincidence of R&D investment and the introduction of an innovation. Even when firms can time the introduction of an innovation separately from R&D investments, higher rates of product obsolescence provide an incentive to match R&D investments with economic booms as product obsolescence starts from the moment of invention.

However, Fabrizio and Tsolmon (2014) perform a number of questionable proceedings when investigating the procyclicality of R&D investments. They for example use an erroneous first-differenced model of R&D investment to examine the timing of R&D investment and it can be questioned whether the variables they introduce to measure product obsolescence and patent effectiveness within an industry are a correct reflection of reality. Furthermore, some of the data manipulations that Fabrizio and Tsolmon

(6)

execute to create a novel data set lead to an inappropriate data set to measure the timing of R&D expenditures.

We will carry out a critical evaluation of the study of Fabrizio and Tsolmon (2014) in which we investigate their model and estimation results and take a closer look at the construction of both the data set and variables. We will conclude with adjusting the estimation sample and introduce a model that works with a selected sample to control for some of the issues that arise from the composition of the data.

The remainder of this paper is organized as follows. The second chapter will further elaborate on the timing of R&D investment by citing and comparing different studies on this subject. In the third chapter we will introduce and critically examine the R&D investment model of Fabrizio and Tsolmon (2014) and provides an adjusted model. The data set that is used for the empirical study together with the construction of the vari-ables are presented in chapter four. Hereafter we will replicate and discuss the results of the R&D investment model in chapter five. Moreover, the fifth chapter will also share the results of a corrected R&D investment model. Chapter six will analyze the robustness of the model by re-estimating the model on different time periods. Furthermore, chap-ter six estimate the R&D investment model on different sub-samples of the data set to provide further insights in the timing of R&D. In the seventh chapter we will introduce a new model that works with a selected sample to mitigate some of the issues that arise due to the data set used by Fabrizio and Tsolmon. Chapter eight presents the conclusion in which the findings of our paper are summarized together with recommendations for further research.

(7)

Chapter 2

Literature Review

The timing of R&D expenditures is a frequently studied topic within economic literature and theories on this subject are often related to research into the relation between the economic state of affairs and productivity growth. One of the prevailing theories on this latter subject comes from Schumpeter (1939) who appoints a positive relation be-tween economic recessions and productivity growth. Because consumer demand declines during a recession, output will decrease and therefore production related activities will generate less profits during this period. Due to this, the opportunity costs of allocat-ing capital towards productivity-related activities and away from production will be less during recessions (Schumpeter, 1939). As productivity enhancing activities have lower opportunity costs during recessions it is expected that firms will invest more time, capital and energy in these activities during a recession. Examples of productivity enhancing activities are staff training, implementing new logistics, rethinking the firms marketing plan and R&D (Saint-Paul, 1993).

Thus, according to the theory of Schumpeter (1939) economic downfalls increase the productivity of firms, which implies that investments in R&D should increase during recessions as R&D is a productivity enhancing activity. Moreover, Barlevy (2005) men-tions that a recession does not decrease the innovative abilities of R&D managers and in combination with the lack of demand for current products it would therefore be an ideal time to allocate capital towards research into and the development of new products and processes. He bases his argument on the paper of Griliches (1990) who found that recessions do not decrease the number of patent applications relative to the research activity.

However, recent empirical evidence shows that R&D expenditures do not grow dur-ing recessions but instead decrease in this period and expand durdur-ing economic booms. One paper that looks further into the timing of innovative activities is that of Geroski and Walters (1995). They show that in the United Kingdom innovative activities like R&D are often clustered around economic expansions and furthermore, they find proof that innovative activity is related to the levels of demand and fluctuates procyclical. Fa-tas (2000) reports this same procyclical pattern when investigating the relation between R&D expenditures and GDP growth in the United States from 1961 to 1996. Another paper that provides evidence for the procyclical pattern in R&D in the US is that of Comin and Gertler (2006). They re-examine the business cycle and construct an alter-native cycle that they call the medium term business cycle. The medium term business cycle continues over a longer period of time, between eight and fifty years, then the ac-cepted business cycle that on average lasts for six year. Using US time series data Camin and Gertler (2006) found evidence that R&D also moves procyclically when implement-ing medium term business cycles. That R&D expenditures exhibit a procyclical pattern in more countries than the US and UK is shown by Wälde and Woitek (2004), who argue that the R&D expenditures of G7 countries between 1973 and 2002 are procyclical.

The reason why R&D expenditures do not grow during recessions might be the 3

(8)

financial constraints that firms face in this period (Aghion et al. 2010). This does not refute the theory of Schumpeter (1939), as firms would still want to invest in R&D during recessions because of the lower opportunity costs, but the recession also imposes credit constraints on firms making it more difficult to invest in R&D during this period. Barlevy (2005) explains the effect of the financial constraints by emphasizing that in order to execute R&D an initial investment is needed, which will pay off in the future. Since demand decreases during a recession firms will have less liquid assets and would therefore need to find funding for their R&D investments. However, Barlevy (2005) mentions that the costs of borrowing money will rise during a recession as banks are also affected by the economic downturn and argues that this affects R&D expenditures in two ways. Firstly firms will refrain from investing in R&D during a recession due to the aforementioned credit constraints and the increased borrowing costs. Secondly, Barlevy (2005) addresses the effect that an economic downturn has on R&D investments that were initiated before the start of the recession. These projects might need to be abandoned as firms are no longer able to secure the funds they needed to implement R&D. This relation is demonstrated by Aghion et al. (2012) who use firm-level panel data from France from 1993-2004 to analyze the effects of credit constraints on R&D expenditures. They report that R&D investment behaves countercyclical when credit constraints are not present, but adopts a procyclical pattern as credit constraints arise. Moreover, Aghion et al. (2012) found that this effect is even more profound for firms in industries that rely more heavily on external financing.

If the theory of Schumpeter (1939) is indeed correct, this would imply that firms that face small or no credit constraints will time their R&D expenditures during recessions. Nonetheless, Barlevy (2004) reports that the R&D growth rate of firms that had at least 50 million dollar in cash or had a net worth of 150 million dollar in 1996 still exhibits procyclical behaviour. Moreover, the R&D expenditures of these rather unconstrained firms exhibit an even more procyclical pattern than the more constraint firms in the

sample. Thus Barlevy (2004) argues that the procyclical behaviour cannot be fully

explained by credit constraints.

As credit constraints do not fully account for the procyclical pattern of R&D, Barlevy (2004) introduces an alternative explanation to why the Schumpeterian opportunity cost hypothesis does not hold for innovative activities like R&D. He argues that due to imitative rivalry, firms delay their R&D expenditures until economic expansions. The idea behind this is that firms want to maximize the profits they make from a new invention and therefore have to take into account the possibility that rivalry firms are able to imitate the new invention and thereby extract a part of the profit. Thus, the profits a firm obtains from a successful imitation will decrease quickly due to spillover effects. Therefore firms will concentrate innovative activity like R&D in economic booms such that the short period in which they make monopoly profits, coincides with times of higher demand.

It is of importance to note that this does not imply that Schumpeter’s (1939) theory is completely rejected. Productivity enhancing activities of which near future profits are not threatened by imitation, will be concentrated in recessionary periods (Barlevy, 2005). This is documented by Nickell et al. (2001), who provide evidence that firms con-centrate productivity enhancing managerial innovations like reorganization or training in economic downturns.

An ancillary element of the alternative explanation of Barlevy (2004) is that it co-incides the investment in R&D with introducing a new innovation, as investing in R&D alone does not directly inform other firms about the new invention. Only after an

(9)

inno-CHAPTER 2. LITERATURE REVIEW 5 vation is introduced on the market or when a patent is filed, other firms can learn from this new idea. Thus, when R&D and introducing a new innovation are contemporaneous then R&D expenditures will behave procyclical due to imitative threat. However, when the introduction of an invention can be delayed then there would be no incentive for firms to time R&D during booms as investigated by Shleifer (1986) and François and Lloyd-Ellis (2003), which is in line with Schumpeter’s (1939) theory. However, empirical evidence documents that firms mostly do not delay the introduction of new inventions. To substantiate his view on the coincidence of R&D and the introduction of a new inno-vation Barlevy (2005) quotes the paper of Griliches (1990), who finds that patents are filed in the beginning of the research process.

Summarizing, there are two explanations on why R&D expenditures do not have the countercyclical pattern as proposed by Schumpeter (1939). Firstly firms face financial constraints during recessions, which makes it hard for firms to find or keep funding their R&D projects. Secondly wanting to maximize profits that originate from the new invention together with the threat of imitation causes firms to time R&D investment in economic booms.

A paper that builds further on the assumption that R&D investments are procyclical because firms focus on profit maximization, is that of Fabrizio and Tsolmon (2014). They empirically test whether the R&D expenditures of firms in industries with weaker intellectual property protection, which therefore are susceptible to imitation, behave more procyclical when compared to firms in other industries. Moreover, Fabrizio and Tsolmon (2014) also investigate the influence of product obsolescence on the timing of R&D expenditures. Just as imitation threat, higher rates of product obsolescence causes profits to decrease after a relatively short period of time and therefore Fabrizio and Tsolmon (2014) assume that higher rates obsolescence will increase the procyclical behaviour of R&D expenditures.

Fabrizio and Tsolmon (2014) thus introduce two hypotheses regarding the timing of R&D expenditures. The first hypothesis is that the R&D investment of firms will be more procyclical in industries with weaker patent protection. Their second hypothesis is that in industries with higher rates of product obsolescence the R&D investments of firms will be more procyclical. If the first hypothesis is supported then imitation threat makes firms time R&D expenditures during booms, which indicates that R&D and innovation are expected to take place simultaneously as proposed by Barlevy (2004). A rejection of the first hypothesis indicates that the strength of patent protection and therefore the protection against imitation does not affect timing of R&D expenditures. In this case it is assumed that investing in R&D and introducing an innovation are two separate processes as predicted by Shleifer (1986) and François and Lloyd-Ellis (2003).

The rate of product obsolescence, which is the main subject of the second hypothesis of Fabrizio and Tsolmon (2014) also makes the profits of a new invention to decrease quickly. There nevertheless is an important difference between product obsolescence and imitation. If R&D and innovation can be timed separately, imitation can be avoided until the firm decides to introduce the product on the market. Product obsolescence however affects the profitability of the innovation as soon as it is invented and regardless of whether or not the invention is introduced on the market. Fabrizio and Tsolmon (2014) therefore argue that faster rates of obsolescence will give firms the incentive to introduce the innovation as quickly as possible on the market to make sure that they reap all the possible profits from the invention. In order to maximize these immediate profits, firms will want to introduce the innovation in periods of high demands and therefore strategically time R&D investments in booms. Timing R&D investments in

(10)

booms enables firms to directly introduce new inventions on the market while demand is high.

Concluding, Fabrizio and Tsolmon (2014) predict that R&D expenditures of firms that are part of an industry with faster rates of product obsolescence will delay investing in R&D until times of higher demand. Furthermore, they assume that firms in industries with lower rates of obsolescence will not have this incentive and will therefore follow the Schumpeterian opportunity cost hypothesis and plan innovative activities like R&D during a bust and delay the introduction of innovations until booms, provided that the timing of R&D and innovation is separate.

The results of Fabrizio and Tsolmon (2014) are consistent with their second hypoth-esis about the effect of product obsolescence on the timing of R&D. They find that the R&D expenditures of firms that are part of an industry with higher rates of obsolescence behave more procyclical. However, Fabrizio and Tsolmon (2014) report that there is no significant proof for the influence of patent effectiveness on the timing of R&D expendi-tures. Thus, their first hypothesis is rejected and the timing of R&D is not affected by imitative threat, which indicates that R&D and the introduction of an innovation are two separate processes.

(11)

Chapter 3

Model

Fabrizio and Tsolmon (2014) introduce two models to estimate the effect of the business cycle on innovative activities of firms. Firstly, they create a model to estimate the effect of changes in industry output on the production of patented inventions. Secondly, they built a model to investigate the effects of these same changes in industry output on R&D investments. We will focus on the second model in order to critically examine the procyclicality of R&D investment that is investigated and affirmed by Fabrizio and Tsolmon. For the R&D investment model Fabrizio and Tsolmon design two different equations. The first estimates the effect of the business cycle on the timing of R&D investments. The second equation can be seen as an expansion of the first by adding the effectiveness of patent protection and the rate of product obsolescence to the model in order to test their additional hypotheses about the procyclicality of R&D investments.

In the following section we will present the R&D investment models with its equations as introduced by Fabrizio and Tsolmon. Hereafter we will provide further insight in the model specifications.

3.1 R&D Investment

In order to investigate the timing of R&D expenditures an explanatory variable that is a representation of the business cycle should be included in the model. However, there are more factors that influence the timing of R&D investments and Fabrizio and Tsolmon use the study of Barlevy (2007) as a guideline for the additional variables they include in their R&D investment model.

Following Barlevy (2007), Fabrizio and Tsolmon include variables that are related to the balance sheets of firms. By doing so they account for the influence that the financial position of a firm can exert on the magnitude of R&D expenditures and its timing. One example Barlevy (2007) gives of the possible effects of firm financials on R&D investments is the relation between firm earnings and the business cycle. Firm earnings are most likely to drop during recessionary periods and therefore there are less financial means to invest in R&D in economic downturns (Barlevy, 2007).

In addition to the variables that are related to the firm’s balance sheet, which we will call the firm-level characteristics, and a one-period-lagged vector of these firm-level characteristics, Fabrizio and Tsolmon follow Barlevy (2007) and add industry output as a regressor to represents the business cycle. To account for time effects, year indicator variables are included in the model. For reasons that will be explained in section 3.2, Fabrizio and Tsolmon implement a first-differenced model, they estimate the growth in R&D expenditures for firm k in year t based on changes in both firm-level characteristics and industry output:

∆RDkt= β0+ β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+

X

τt+ ωkt (3.1)

(12)

∆Mkt and the lagged version ∆Mkt−1 represent vectors of firm-level characteristics for

firm k in respectively year t and t − 1. These vectors consist of different independent variables that together capture the financial position of a firm. Section 4.2 will further

elaborate on the composition of these vectors. ∆Xit represents changes in industry

output for industry i in year t. Because changes in output that are captured by ∆X_it

are expected to be correlated across industries due to business cycle effects, Fabrizio and

Tsolmon include year dummies τ_t to account for time effects.

As the firm-level variables are included to control for other events that are likely to

have an influence on R&D expenditures, β3 captures the relation between changes in

industry demand and the R&D expenditures of firms. A larger than zero and significant

coefficient β₃ implies that firms increase their R&D investments when industry output

grows and therefore indicates a procyclical nature. However, a significant and smaller

than zero coefficient β3 indicates a countercyclical pattern R&D investments. In this

case firms increase their R&D expenditures when industry output declines.

To test their hypotheses about the impact of product obsolescence and the effec-tiveness of patent protection on the timing of R&D expenditures, Fabrizio and Tsolmon introduce a second R&D investment equation where they interact abovementioned vari-ables with the industry output variable. This brings us to the second model:

∆RDkt= β0+ β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+ β4∆Xit× Obs

+ β5∆Xit× P atef f +

X

τt+ ωkt

(3.2)

By introducing the interaction terms ∆Xit × Obs and ∆Xit × P atef f Fabrizio and

Tsolmon are able to investigate the impact of obsolescence and patent effectiveness on

the timing of R&D expenditures. A positive and significant β₄ indicates that firms that

are part of industries with faster rates of obsolescence will have a larger increase in R&D expenditures when industry output grows compared to firms in industries with smaller

rates of obsolescence. That is to say, a significant and positive coefficient β₄suggests that

R&D expenditures of firms that are part of industries with higher rates of obsolescence have a more procyclical pattern.

The coefficient β₅ gives insight in the relation between the effectiveness of patent

protection and the timing R&D expenditures. A smaller than zero and significant co-efficient would suggest that firms that are part of industries with more effective patent protection are less sensitive to changes in industry output. Therefore the change in R&D expenditures as a reaction to changes in industry output will be less for firms that are part of industries with stronger patent protection compared to firms in industries with weaker patent protection.

While Fabrizio and Tsolmon (2014) only mention the model specifications as in equa-tion 3.1 and 3.2 in the empirical methodology of their paper, the empirical results secequa-tion shows that they estimated another model that is referred to as the full model:

+ β5∆Xit× P atef f + β6∆Xit× ExF in +

X

τt+ ωkt

(3.3) The full model includes the reliance on external finance per industry group and interacts it with the industry output. Fabrizio and Tsolmon add this interaction in order to take into account that industries that more heavily rely on external finances, will face heavier financial constraints during recessionary periods. A positive and significant coefficient

β6 suggests that the R&D expenditures of firms that are part of an industry that relies

(13)

CHAPTER 3. MODEL 9 Something else that stands out in the estimation results of Fabrizio and Tsolmon is the addition of time invariant regressors to the model while these are not included in equation 3.2. Thus, the full model estimated by Fabrizio and Tsolmon is actually as follows:

+ β5Obsi+ β6∆Xit× P atef f + β7P atef fi

+ β8∆Xit× ExF in + β9ExF ini+

X

τt+ ωkt

(3.4)

We will discuss the implications of adding time invariant regressors to the model when specifying the first-differenced model in section 3.2.2 and when evaluating the R&D replication results in section 5.1.1.

3.2 Model Specification

Fabrizio and Tsolmon (2014) use a panel that observes the same cross section of 7,731 firms during a time period from 1975 until 2002. They deal with an unbalanced panel as due to missing values not all firms in the data set are observed for each year between 1975 and 2002. That is to say, the firms in the data set are observed at different moments in time and therefore the number of observations can differ per firm.

3.2.1 Generic Panel Data Model

A standard panel data regression model with i = 1, ..., N different cross-sectional units that are observed during time t = 1, ..., T is written as follows:

Yit= α + Xit0β + Z

0 iρ + it

it= µi+ uit+ λt

(3.5)

where Yit represents the dependent variable that is estimated by a constant α, a K

dimensional vector of explanatory variables Xit, observed time invariant individual

char-acteristics Z_i and the error term _it. λ_t is the time variant part of the error term, which

causes the intercept of the error term to differ over years. Including dummy variables

for each year can control for these time-specific effects. The µi variable in the error term

models unobserved, time invariant individual characteristics.

The OLS estimation of the parameters in equation 3.5 is consistent when we assume

that µi is not correlated with the explanatory variables Xit of the model. If there is no

correlation between µ_iand X_itthen any unobserved individual heterogeneity that might

appear due to the individual effects µi is independently distributed over the regressors

(Cameron and Trivedi, 2005). In this case the effects µi are treated as random effects.

However, given the observational nature of our data it is more reasonable to expect that the time invariant effects in the error term are correlated with the explanatory vari-ables of the model. Examples of time invariant varivari-ables that can affect the explanatory variables of the model are for instance the geographic location, corporate structure or

business culture of the firm. Therefore we assume the effects µi to be fixed, which causes

the OLS regression to suffer from an endogeneity problem, thus the OLS estimates of the model in equation 3.5 will be biased and inconsistent.

As mentioned in section 3.1, Fabrizio and Tsolmon do not use the generic panel data model with level variables, but instead use the first-differenced model of R&D investment

(14)

of Barlevy (2007). However, the reasons they give for doing so are inconclusive. We would expect that they made the choice based on the abovementioned possible presence of fixed effects that cause the OLS estimator to be biased. Moreover, Barlevy (2007) indeed regresses the growth of R&D expenditures on industry output growth, using firm fixed effects thereby indicating the presence of fixed effects. Nevertheless Fabrizio and Tsolmon do not test for the presence of fixed effects, which could be done by a Hausman specification test, but rather test for autocorrelation and heteroskedasticity while the occurrence of both does not by itself effect the unbiasedness of the OLS estimator.

Fabrizio and Tsolmon find proof for autocorrelation and heteroskedasticity when test-ing the the R&D investment model ustest-ing level variables. Fabrizio and Tsolmon state that a Wooldridge test for autocorrelation resulted in F (1, 6823) = 1, 858.92, which cor-responds to a p-value of 0.000 and thus rejects the null hypotheses of no autocorrelation. Furthermore, they mention that the Breusch-Pagan test produced a chi-square statistic of 6,355.65 with a corresponding p-value of 0.000 and thus rejects the null hypotheses that the variance of the error terms is constant.

While the abovementioned results on heteroskedasticity and autocorrelation indicate that the OLS estimators of the R&D investment model when estimated as in equation 3.5, are not the Best Linear Unbiased Estimators (BLUE), it does not directly indicate

the presence of fixed effects. Nevertheless, instead of running an additional test to

test for fixed effects Fabrizio and Tsolmon state that they use a first-differenced model to account for the autocorrelation and heteroskedasticity. While both only indicate a problem if they are related to fixed or random effects within the model.

After first-differencing the model, Fabrizio and Tsolmon ran a Durbin-Watson test and found no remaining autocorrelation. We argue that first-differencing the model

diminishes the problems because they emerged from the fixed effects in the model1. The

following section will further discuss the transformation of the general panel model into a first-differenced model and shows how taking first-differences accounts for fixed effects.

3.2.2 First-Differenced Model

As seen in section 3.2.1, the generic panel data model can contain unobserved, time

invariant individual characteristics µi that are correlated with the independent variables

of the model. First-differencing the model is a way to control for the omitted variable

bias that occurs due to the fixed effects µ_i.

To obtain the first-difference estimator the one-period lagged observation needs to be subtracted from the original equation as portrayed in 3.5. This leads to the following model:

Yit− Yit−1 = (α − α) + (Xit0 − Xit−10 )β + (Zi0− Zi0)ρ + (it− it−1)

it− it−1 = (µi− µi) + (uit− uit−1) + (λt− λt−1)

(3.6) and rewriting this equation results in:

∆Yit= ∆Xit0 β + ∆uit+ ∆λt (3.7)

1_{Testing the R&D investment model for fixed effects, we run a Hausman test on the level variable}

variant of the models as specified in equation 3.1 and 3.2. Both Hausman tests resulted in a chi-square statistic with a p-value of 0.000. This rejects the null hypothesis that the random effects model is consistent.

(15)

CHAPTER 3. MODEL 11 The first-difference estimator thus specifies the relation between one-period changes in the explanatory variables on individual level and one-period changes in the dependent

variable. First-differencing the model removes µi, the constant α and the time invariant

individual characteristics Z_i. Moreover, a consequence of taking first-differences is that

one time observation is lost.

Comparing the analytical results of the first-difference model with the model of Fab-rizio and Tsolmon as denoted in equation 3.1, 3.2 and 3.3 it becomes apparent that Fabrizio and Tsolmon did not remove the constant term from their model equations.

This is justified when one of the time dummies τtis dropped from the equation, however

Fabrizio and Tsolmon include dummy variables for each year and therefore adding a constant term is redundant. Moreover, as depicted in equation 3.4, Fabrizio and Tsol-mon include time invariant industry characteristics in their model, which makes their first-differenced models incorrect. Correct first-differenced models for R&D investment are: ∆RDkt= β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+ X τt+ ωkt (3.8) ∆RDkt= β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+ β4∆Xit× Obs + β5∆Xit× P atef f + X τt+ ωkt (3.9) ∆RDkt= β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+ β4∆Xit× Obs + β5∆Xit× P atef f + β6∆Xit× ExF in + X τt+ ωkt (3.10) Fabrizio and Tsolmon give no explanation on why they kept the constant term or time invariant variables in their first-differenced model. We are of opinion that it makes no sense to include time invariant industry characteristics in the first-differenced model as theory clearly states that these should drop out. There can be different reasons to keep the constant term in a regression model, it ensures that the model residuals have a mean of zero and dropping the constant forces the regression line to go to the origin.

However, the intercept of a first-differenced model can be equal to zero as the model estimates the linear relation between changes in value instead of the relation between the real values of the variables. Referring to our model, we expect that amount invested in R&D will stay the same for two subsequent years when the value of all other variables in our model also stays constant for those years. Thus, when there are no changes in the values of the regressors from year t-1 to t, we can assume that the R&D expenditures

will also stay the same and ∆RDkt will be equal to zero. Therefore we do not consider

the zero intercept to be a problem for the R&D investment model.

It is important to note that this only holds when there is no linear trend in R&D expenditures. When R&D expenditures exhibit a linear trend δ · t, the generic panel data and first-differenced model will be:

Yit= α + Xit0β + Zi0ρ + it

it= µi+ uit+ λt+ δ · t

(3.11)

such that

(16)

It is therefore necessary to include a constant in a first-differenced model when it

is expected that there is a linear trend in the dependent variable. However, we do

not assume that R&D investment displays trending behaviour as the model is already corrected for inflation. There could be small factors influencing the R&D expenditures of all firms, like policy changes or new tax rules that make investing in R&D more worthwhile but we believe that this are temporary effects and do not cause an upward of downward trend in R&D.

Summarizing the abovementioned arguments, we argue that there is no valid reason to keep the constant in the R&D investment model. The model already includes dummies for every year to capture time fixed effects and the model is not expected to exhibit a linear trend thus its intercept can be equal to zero. Therefore we will estimate the adjusted versions of the first-differenced R&D investment models without the constant or time invariant variables in chapter 5 and elaborate on the regression results.

Except for the fact that Fabrizio and Tsolmon made some errors while composing their first-differenced model of R&D investment, we assume that the problems that occurred due to the fixed effects are solved by first-differencing the model. Nevertheless, to be sure that the first-differenced models controls for heteroskedasticity, Fabrizio and Tsolmon use clustered standard errors. Clustering the standard errors on firm level is justified since we assume that the yearly outcomes of the variables for the same firm are correlated.

Another potential problem are the industry groups. The R&D expenditures of firms depend on the industry output, which is equal for firms that fall within the same indus-try thus, the R&D expenditures for firms within the same indusindus-try are not independent. Moreover, higher competition in certain industry groups can for instance influence the R&D expenditures of firms in these industries. However, we assume that Fabrizio and Tsolmon partly control for this by adding industry output, patent effectiveness, obsoles-cence and external finance to their model. It would be of interest for further research to see how the timing of R&D expenditures differs between industry groups and how to account for this within the model. This subject will be slightly touched on in section 6.1.

Concluding, Fabrizio and Tsolmon introduce a first-differenced model of R&D in-vestment with standard errors clustered by firm. First-differencing the model accounts for the problems that arise due to the presence of fixed effects and the inclusion of year dummies controls for the time-specific effects.

(17)

Chapter 4

Data and Variables

We use the replication data set from the Economic Review of Statistics to get more insight in the R&D investment model of Fabrizio and Tsolmon (2014). The data set encloses the different data sources together with different programming files made by Fabrizio and Tsolmon.

Using the replication data, this chapter will discuss the different data sources that Fabrizio and Tsolmon use to create a novel data set for their research. It will explain both the origin and content of the data sources, together with possible disadvantages that arise due to the design of the data sources. Moreover, the chapter provides further information about the variables of the R&D investment model and explains how they are constructed from the different data sources. The chapter concludes with some descriptive statistics.

4.1 Data

Fabrizio and Tsolmon combine three different data sources in order to create an unique data set that is used to estimate their R&D investment model. First, the Compustat database from 1975-2002 is included to derive key information on firms over this period. Because the Compustat database contains financial data and market information on a large range of publicly traded global companies, it is possible to obtain different firm-year financials together with information about external financing per industry group. More-over, one of the key aspects of the Compustat database is that it contains information about the yearly R&D expenditures per firm.

Secondly, the NBER Manufacturing and Productivity database is used in order to obtain information about changes in industry output. The NBER Manufacturing and Productivity database is developed by Bartelsman and Grey (1996) and is publically ac-cessible via their website. It contains data about the industry level demand per firm and as the Compustat database it links firms to the industry group in which they operate, which makes it possible to match both databases and relate the R&D expenditures per firm to changes industry output. Furthermore, the NBER Manufacturing and Produc-tivity database is consulted to obtain annual GDP deflator data that will be used to deflate some of the explanatory variables from the Compustat database.

The third data source is the Carnegie and Mellon survey (CMS) of R&D managers (Cohen et al., 2000), which is used to access data on the rates of obsolescence and the effectiveness of patent protection within an industry. The CMS survey questioned R&D managers from 240 different R&D labs that conducted R&D within the manufacturing

industry in the United States. The next section will give more details on how the

variables for obsolescence and patent effectiveness are constructed from the CMS survey responses.

(18)

The CMS survey has certain limitations. The survey is administered once and there-fore only has information about the patent protection effectiveness and obsolescence within industries at one moment in time, the 1991 to 1993 period. The consequence hereof is that the data set of Fabrizio and Tsolmon has a constant value for both vari-ables for the period 1975-2002, while the effectiveness of patent protections and the obsolescence within industries are expected to change over the years. For further re-search it would therefore be an improvement to obtain information about these industry level characteristics on a more regular basis.

4.2 Construction of the Variables

The variables of the R&D investment model can be placed in two categories. The first category includes the variables of interest that are used to investigate the timing of R&D investment and the effect of both imitation threat and product obsolescence on this timing. Secondly, a group of control variables that represent firm-level characteristics are included in order to account for the financial position of firms, which can possibly have an influence on the timing of R&D investment.

Table 4.1 presents an overview of all variables together with their data source and the level on which they operate. The remaining part of this section will provide a short explanation on how the variables of both abovementioned categories are constructed.

Table 4.1: Description of the Variables

Variables Description Level Database Dependent Variable

∆RDkt Delta_R&D First difference of the natural log of R&D expenditures per year Firm Compustat

Explanatory Variables

∆Mkt Delta_Cash Flow First difference of the natural log of firm’s cash flow per year Firm Compustat

Delta_Total Assets First difference of the natural log of firm’s total assets per year Firm Compustat Delta_Total Liabilities First difference of the natural log of firm’s total liabilities per year Firm Compustat Delta_LT Debt First difference of the natural log of firm’s long-term debt per year Firm Compustat Delta_ST Debt First difference of the natural log of firm’s short-term debt per year Firm Compustat Delta_Capital Stock First difference of the natural log of firm’s net property, plant and equipment per year Firm Compustat ∆Mkt−1 Delta_Cash Flow_Lag One period lagged first difference of the natural log of firm’s cash flow per year Firm Compustat Delta_Total Assets_Lag One period lagged first difference of the natural log of firm’s total assets per year Firm Compustat Delta_Total Liabilities_Lag One period lagged first difference of the natural log of firm’s total liabilities per year Firm Compustat Delta_LT Debt_Lag One period lagged first difference of the natural log of firm’s long-term debt per year Firm Compustat Delta_ST Debt_Lag One period lagged first difference of the natural log of firm’s short-term debt per year Firm Compustat Delta_Capital Stock_Lag One period lagged first difference of the natural log of firm’s net property, plant and equipment per year Firm Compustat ∆Xit Delta_Output First difference of the natural log of industry real gross output per year Industry NBER ∆Xit× Obs Output × Obsolescence Industry NBER - CMS ∆Xit× P atef f Output × Pattent Effectiveness Industry NBER - CMS ∆Xit× ExtF in Output × External Financing Industry NBER - Compustat

Pattent Effectiveness Average survey response to degree of effectiveness of patent protection, constant over years Industry CMS Survey Obsolescence Average survey response to rate of introduction of innovations, constant over years Industry CMS Survey External Financing Extent of industry reliance on external financing, constant over years Industry Compustat

(19)

CHAPTER 4. DATA AND VARIABLES 15

4.2.1 R&D investment growth

The dependent variable of the model originates from the R&D variable in the Compustat data set that expresses the yearly R&D expenditures for each firm. After setting all missing values for R&D within the data set equal to zero, Fabrizio and Tsolmon deflate the yearly R&D expenditures with the GDP deflator from the NBER Manufacturing and Productivity database. Taking first-differences of the natural log of the deflated yearly R&D expenditures creates the eventual R&D growth variable.

Figure 4.1 provides insight into the weighted average growth of R&D expenditures for firms within the Compustat database. Imitating Barlevy (2007), we take all firms in the Compustat database that reported positive R&D in both year t and year t − 1. We calculate the difference in reported R&D for these subsequent years and divide the difference by the R&D expenditures in year t to obtain the weighted average R&D growth. After taking the log of the weighted average R&D growth, we followed Barlevy (2007) and minimize the effect of outliers by removing observations where the absolute value of log growth is above the 95th percentile. The resulting weighted average log growth in R&D expenditures is shown in figure 4.1.

Figure 4.1: Weighted average growth of R&D expenditures

Figure 4.1 shows a strong decline in R&D growth in the early seventies, which is followed by almost five years of growth that ended at the beginning of the eighties. The nineteen-eighties are marked by an overall decline in R&D growth, which reflects the multiple crises that occurred in this period. The long period of growth in the nineteen-nineties however manages to restore the R&D growth to roughly the same level as in the beginning of the eighties. The peak at the turn of the century is followed by a sharp decline, which marks the end of the economic expansion of the nineties and the beginning of the early two-thousands recession (Hall et al. 2001a).

(20)

4.2.2 Firm level controls

The group of control variables consist of six different variables together with a lagged version of all these variables, which represent the firm-level characteristics. The first-difference of the natural logarithm of the following variables together with a lagged version are added as firm level controls: cash flow, total assets, total liabilities, long-term debt, short-long-term debt, capital stock. All variables are obtained from the Compustat database.

4.2.3 Industry output growth

Fabrizio and Tsolmon create a nominal gross output variable by adding up the data on annual value added and materials costs for each industry group in the NBER Manufac-turing and Productivity database. Afterwards, they deflate the nominal gross output using the shipments deflator of the corresponding industry, which is also provided in the NBER Manufacturing and Productivity database. This results in the annual real gross output per industry group. It is possible to link the annual real gross output per industry group to the firms in the general data set because the latter data set has information about the industry group in which a firm is located. After merging the data sets, the growth of industry output is estimated by taking the first-difference of the natural log of the annual real gross output output variable.

4.2.4 Rate of obsolescence

The rate of obsolescence within each industry group is obtained from the CMS survey. In the survey, R&D managers were asked to indicate the speed whereby new products and processes are introduced within their main industry. The answers had to be given on a scale of one to five, one indicating very slow process and product innovation and five being very rapidly. Using the survey outcomes, Fabrizio and Tsolmon create the weighted averaged introduction speed for every firm by dividing the introduction speed by the percentage of R&D investment that went into product and process innovation. Because it is known of which industry a firm is part of, the eventual average rate of obsolescence variable is obtained by calculating the average speed of product and process introduction per industry group.

A disadvantage of using the average speed per industry group instead of on firm level is that the survey responses for firms that fall within the same industry group can vary, such that a firm with slow process and product innovation can be categorized as having a rapid rate of obsolescence. Fabrizio and Tsolmon look at the response dispersion and concluded that even as the answers per firm within an industry vary, the vast majority of the responses match the economic literature on obsolescence rates in industries.

The industry groups in the data set are encoded with a Standard Industrial Clas-sification (SIC) code which enables us to link observations to a certain industry group or major industry group. To analyze the dispersion of obsolescence rates over the sam-ple, we calculate the average introduction speed of new product and process innovations per major industry group based on the SIC codes within the sample. The results are displayed in figure 4.2. Section 6.1 will discuss the composition and structure of the industries and the SIC codes more extensively. Table A.1 in appendix A gives a detailed description of the major industry groups that are represented by the SIC codes.

(21)

Figure 4.2: Average obsolescence per SIC code

Figure 4.2 displays that the average rate of obsolescence is faster than moderate in major industry groups 21, 24, 27 and 31. These industries are related to tobacco products, lumber and wood, printing and publishing and leather goods.

4.2.5 Effectiveness of patent protection

The patent effectiveness variable is also based on the CMS survey results. The CMS survey asked R&D managers to indicate for which percentage of the product and process innovations that were patented, the patent was effective. With an effective patent being a patent that makes sure that other rivalry firms cannot imitate the new product, ensuring the firm to have a monopoly on the new invention and therefore be able to capture maximal rents. Fabrizio and Tsolmon follow Arora and Ceccagnoli (2006) and calculate the average of the product and process patent effectiveness weighted by the percentage of R&D expenditures that is allocated to product and process innovation. After calculating the weighted average patent score for every firm, Fabrizio and Tsolmon averaged these scores per industry group to obtain the patent effectiveness measure per industry group. Figure 4.3 shows the average patent effectiveness per major industry group from 1975 to 2002. Table A.1 in Appendix A provides a more detailed description of the major industry groups.

(22)

Figure 4.3: Average patent effectiveness per SIC code

Figure 4.3 displays that for firms in industry groups 25, 28 and 38 the average rate of patent effectiveness is more than fifty percent, indicating that in these industry groups more than half of the filed patents were effective. These industries are related to furniture and fixtures, chemicals and allied products and different kinds of mechanical instruments. For industry groups 21, 27 and 24, which represent industries related to tobacco products, printing, publishing and lumber and wood products the patent effectiveness was lowest, ranging from five to eighteen percent.

4.2.6 Reliance on external financing

While Fabrizio and Tsolmon do not include the reliance on external finance as a variable in the model as specified in equation 3.2, they do control for this factor when estimating the R&D investment model. Information on the amount of external funds that firms use to support their investments is obtained from the Compustat database. Dependence on external funds is calculated as a ratio between the aforementioned amount of external finance and the total capital expenditures of the firm. The reliance on external finance is hereafter calculated as an average of the firm-level reliance per industry group. Figure 4.4 shows the average reliance on external finance per major industry group. Positive values for average external finance indicate that the industry group relies on external financing for their investments. Negative values represent the presence of sufficient internal funds to finance investments. From figure 4.4 we can deduce that the industry group that represents chemicals and allied products relies most on external finance and the tobacco industry has the most internal cash available to finance investments. A description of the industry group codes can be found in table A.1 in the appendix.

(23)

Figure 4.4: Average reliance on external finance

4.2.7 Time effects

The year indicator variables are time dummies for 1975 until 2002 that are included in the model to capture any time-specific effects. These time dummies account for any trends that affect all firms in the data set, but are not captured by the included variables. Examples of time trends are the development of and increase in online shopping, the use of big data and increasing consumer awareness.

(24)

4.3 Descriptive Statistics

Table 4.2 reports the descriptive statistics of the variables of the model. As seen from table 4.2, the model is estimated on a data set consisting of 71,264 observations on 7,754 firms from 1975 until 2002. Moreover, table 4.2 provides the descriptive statistics of the natural log of the variables underlying the first-differenced variables of the model.

Table 4.2: Descriptive Statistics Key Variables

N N(firms) Mean Median SD Minimum Maximum Dependent Variable Delta_R&D 71264 7754 0.11 0.00 1.09 -11.69 12.94 Explanatory Variables Delta_Output 71264 7754 0.04 0.04 0.11 -0.53 1.12 Patent Effectiveness 71264 7754 0.44 0.44 0.15 0.05 0.82 Obsolescence 71264 7754 2.84 2.87 0.35 1.56 4.00 External Financing 71264 7754 0.33 -0.14 1.39 -2.93 4.76 Delta_Cash Flow 71264 7754 0.01 0.02 0.51 -8.54 8.30 Delta_Total Assets 71264 7754 0.18 0.03 1.33 -14.64 14.57 Delta_Total Liabilities 71264 7754 0.18 0.02 1.31 -13.93 14.29 Delta_LT Debt 71264 7754 0.12 -0.01 1.74 -12.95 13.64 Delta_ST Debt 71264 7754 0.09 0.00 1.62 -14.00 14.22 Delta_Capital Stock 71264 7754 0.16 0.01 1.17 -13.97 14.00 R&D 71264 7754 3.64 3.94 3.36 0 13.76 Output 71264 7754 8.92 8.72 1.06 5.79 12.37 Cash Flow 71264 7754 7.83 7.32 1.42 -1.66 15.14 Total Assets 71264 7754 8.74 8.80 2.83 0 17.39 Total Liabilities 71264 7754 7.94 7.9 2.86 0 17.37 LT Debt 71264 7754 5.87 6.34 3.76 0 16.37 ST Debt 71264 7754 3.84 4.02 2.88 0 14.80 Capital Stock 71264 7754 7.26 7.39 3.05 0 16.08 Year 1975 2002

The natural log of R&D expenditures has a minimum value of zero and a maximum of 13.76 with a median of 3.64. The growth in R&D expenditures has a minimum of -11.69, which represents a cut in R&D expenditures and a maximum of 12.94. However, the median of R&D expenditures growth is zero, which indicates that the half of the observations on changes in R&D expenditures is equal to or smaller than zero. Taking a closer look at the data, we see that roughly 40% of the observations of R&D expenditures growth are equal to zero. We will further discuss this issue in chapter 7.

Other variables of interest are obsolescence and patent effectiveness. As mentioned before, obsolescence is measured on a scale from 1 to 5, with 1 indicating very slow, 3 moderate and 5 very rapid obsolescence speed. The maximum rate of obsolescence within the data set is 4, indicating rapid obsolescence and the minimum is 1.56, which expresses a rate of obsolescence between very slow and slow. The percentage of effective patents ranges from 5% up to 82% with a median of 44%.

(25)

Chapter 5

Replication of the R&D Model

To perform the replication of the R&D investment model we use the R&D replication Stata do-file obtained from the Economic Review of Statistics. The do-file needs to be adjusted due to problems that arise when the different data sources are combined to create the unique data set for the R&D investment replication. Furthermore, some other small errors occur when running the do-file for which we need to adjust the programming code. We expect that the altered do-file performs in the same way as the program used by Fabrizio and Tsolmon (2014). That is to say, the abovementioned modifications of the do-file should not have any impact on the structure of the model.

This chapter is structured as follows. First the replication and its results will be discussed in section 5.1. Section 5.1.1 will hereafter evaluate the choices Fabrizio and Tsolmon made while estimating the R&D model. The chapter will conclude with the replication of an adjusted R&D investment model. The implemented adjustments ac-count for several errors that Fabrizio and Tsolmon made when defining their model and performing the regression.

5.1 Replication Results

After running a pooled regression on the R&D investment model of Fabrizio and Tsolmon using clustered standard errors, we obtain results over a set of 71,246 observations that consist of 7,754 firms in 102 different manufacturing industries. While the number of observations and industries match the numbers in the paper of Fabrizio and Tsolmon, the number of firms that are taken into account when running the model does not equal their results. It is not clear why our results represent 23 extra firms in comparison with those of Fabrizio and Tsolmon, given that we did not make any profound changes to the content of the model. Moreover, Fabrizio and Tsolmon provide ambiguous results on the number of firms in the R&D investment set. While the data and variables section in their paper states that the data set contains results on 7,731 firms, the introduction of their paper mentions a panel data set of 7,754 public firms, which corresponds with our results.

Fabrizio and Tsolmon also mention the number of firms in their data set before the sample is matched to the CMS survey data. We compare whether our sample size matches that of Fabrizio and Tsolmon before the CMS survey in order to examine where the discrepancy between the number of firms in the data arises. Running the R&D model without matching it to the CMS survey data results in a set of 75,093 observations that consists of 8,165 firms in 118 different manufacturing industries. This corresponds with the numbers of Fabrizio and Tsolmon and therefore we conclude that the difference in the number of firms arises when the data is matched to the CMS survey data. Furthermore, the descriptive statistics in table 4.2 are similar to those of Fabrizio and Tsolmon, with the exception of the number of firms. Therefore we argue that Fabrizio and Tsolmon

(26)

most likely made a mistake when writing down the amount of firms in the estimation sample.

Moreover, some other differences in the estimation results are seen when we com-pare the results in table 5.1 with those of Fabrizio and Tsolmon. While the sign and significance of the coefficients are almost always equal to the results of Fabrizio and Tsolmon, there are some discrepancies in the magnitude of the results. It is not clear why these variations in results for the R&D model arise, but we assume that it might be related to the difference in the number of firms that we mentioned earlier. That is to say, matching the data with the CMS survey data seems to effect our replication in such a way that the results are not exactly in line with those in the paper of Fabrizio and Tsolmon. Because we use the data and follow the programming files provided by Fabrizio and Tsolmon, we do not know where the discrepancies that arise during the matching the CMS survey data come from. Nevertheless, the results are approximately the same and Fabrizio and Tsolmon themselves make ambiguous statements about the number of firms in their panel data set. For these reasons we assume that are results are in line with those of Fabrizio and Tsolmon and we will use the data and model as provided for further research in this paper.

Table 5.1 presents the replication results of the R&D investment model. The first column reports the results of the R&D investment model as in equation 3.1 without controlling for common time effects. The positive and significant sign for industry output affirms the procyclicality of R&D investment. The same estimation as in the first column is reported in the second column, but now while controlling for common time effects by adding year dummies to the regression. The estimation results in the second column still affirm the procyclical pattern in R&D investment.

The hypothesis whether R&D investment will be more sensitive to changes in industry output for firms in industries with weaker patent protection, is tested in the third column. The coefficient for the interaction between output and patent effectiveness is larger than zero, which suggests that firms in industries with stronger patent protection behave more procyclical. This refutes the hypothesis of Fabrizio and Tsolmon, but the result is not significant. Therefore we have no indication that the patent effectiveness within an industry effects the procyclical behaviour of the R&D expenditures of firms within this industry.

Column four reports the results of testing the hypothesis that firms in industries with higher rates of obsolescence have a more procyclical pattern of R&D expenditures. This hypothesis is supported as the coefficient of the interaction between output and obsolescence is positive and significant at the 0.1% level.

The results of the R&D investment model as in equation 3.2 are reported in the fifth column. The coefficients and significance of both interaction terms do not change much in respect to the third and fourth column. The sixth column includes a new interaction term to the full model, the interaction between external finance and industry output. Fabrizio and Tsolmon add this interaction term to account for the fact that firms that rely heavily on external financing, face more financial constraints during recessions. However, including the interaction term does not severely impact the results. The same conclusions as in column five still hold.

The seventh and last column estimates the whole model with inclusion of the inter-action between output and external financing, but without the year indicator variables that capture the aggregate business cycle effects. From the seventh column we can de-rive that the time trends do not have a big impact on the regression results as they stay approximately the same.

(27)

CHAPTER 5. REPLICATION OF THE R&D MODEL 23 Based on the estimation results, Fabrizio and Tsolmon conclude that firms that are part of an industry in which products have higher rates of obsolescence will match their R&D investment with economic booms in order to maximize profits. Furthermore, the results provide no evidence of a relation between patent effectiveness and the timing of R&D expenditures.

Table 5.1: OLS Estimates of R&D Expenditure Growth

(1) (2) (3) (4) (5) (6) (7)

Delta_Output 0.204*** 0.272*** 0.176 -0.854*** -0.777** -0.747* -0.814**

(0.031) (0.034) (0.118) (0.241) (0.262) (0.292) (0.290)

Output × Patent Effectiveness 0.272 -0.054 -0.178 -0.224

(0.299) (0.306) (0.314) (0.304) Pattent Effectiveness 0.172*** 0.183*** 0.101*** 0.100*** (0.023) (0.023) (0.026) (0.026) Output × Obsolescence 0.357*** 0.341*** 0.341*** 0.351*** (0.079) (0.079) (0.088) (0.087) Obsolescence 0.047*** 0.048*** 0.033** 0.036*** (0.009) (0.010) (0.010) (0.010)

Output × External Financing 0.039 0.060

(0.063) (0.061) External Financing 0.014*** 0.014*** (0.004) (0.004) Delta_Cash Flow -0.031* -0.028 -0.028 -0.027 -0.026 -0.025 -0.027 (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) Delta_Cash Flow_Lag -0.006 -0.007 -0.006 -0.006 -0.005 -0.004 -0.003 (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) Delta_Total Assests 0.249*** 0.248*** 0.248*** 0.247*** 0.247*** 0.247*** 0.247*** (0.018) (0.018) (0.018) (0.018) (0.018) (0.018) (0.018) Delta_Total Assests_Lag 0.090*** 0.089*** 0.088*** 0.088*** 0.087*** 0.086*** 0.086*** (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) Delta_Total Liabilities 0.076*** 0.076*** 0.076*** 0.076*** 0.076*** 0.076*** 0.076*** (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) Delta_Total Liabilities_Lag -0.076*** -0.076*** -0.075*** -0.076*** -0.075*** -0.075*** -0.075*** (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) Delta_LT Debt -0.024*** -0.024*** -0.023*** -0.023*** -0.023*** -0.023*** -0.023*** (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Delta_LT Debt_Lag -0.000 0.000 0.000 0.000 0.000 0.001 0.000 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Delta_ST Debt -0.001 -0.001 -0.001 -0.001 -0.001 -0.001 -0.001 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Delta_ST Debt_Lag 0.004 0.004 0.004 0.004 0.004 0.004 0.004 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Delta_Capital Stock 0.104*** 0.103*** 0.103*** 0.103*** 0.103*** 0.103*** 0.104*** (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) Delta_Capital Stock_Lag 0.001 0.001 0.001 0.001 0.001 0.001 0.001 (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) Constant 0.022*** 0.085*** 0.013 -0.105*** -0.188*** -0.110** -0.128*** (0.003) (0.022) (0.024) (0.032) (0.034) (0.037) (0.032)

Year fixed effects No Yes Yes Yes Yes Yes No

N (observations) 71,246 71,246 71,246 71,246 71,246 71,246 71,246

R2 _0.22 _0.22 _0.22 _0.22 _0.23 _0.23 _0.22

(28)

5.1.1 Evaluation of the R&D Estimation

There are some problems with the empirical estimation of the R&D investment model of Fabrizio and Tsolmon. As seen from table 5.1, Fabrizio and Tsolmon include a constant and time-invariant variables for patent effectiveness, obsolescence and external finance in the model while these should not be part of a first-differenced model. Moreover, the addition of the interaction between output and external to the model was not mentioned in the empirical methodology of their paper. The addition of this interaction term can nevertheless be justified since Fabrizio and Tsolmon run two additional regressions to check whether the original model is rightly specified or if it suffers from the effects that the reliance on external financing within an industry can have on the timing of R&D.

We will re-estimate the R&D investment model in section 5.2 to see how the inclusion of the time invariant variables affected the estimation results. Furthermore, this section will estimate the model of Fabrizio and Tsolmon without the inclusion of a constant and examine whether it is necessary to include the constant term even though its significans at the 0.1% level is reported in table 5.1. That is to say, estimating the model without a constant makes it possible to examine whether or not there is a linear trend in R&D expenditures that needs to be accounted for.

The second problem that arises is related to the construction of the data set that is eventually used to estimate the R&D investment model. In the data and variable section of their paper, Fabrizio and Tsolmon state that they only kept observations of firms that did not have missing values for R&D expenditures. Nevertheless, the programming file does not drop firms with missing values of R&D, instead it replaces missing values for R&D expenditures with a value of zero. Before this data manipulation only 2,178 of the observations between 1975 and 2002 had a zero value for R&D expenditures. After the manipulation the data set contained 27,207 observations with zero R&D expenditures causing approximately 40% of the observations on R&D within the estimation sample to be equal to zero. We will correct for the large amount of zeros that resulted from the abovementioned data adjustment in chapter 7.

5.2 Correcting the R&D Estimation

While examining the characteristics of first-differenced models in section 3.2.2 we atgued that the constant term should be excluded from the R&D investment model given that any presumed time effects are captured by the year indicator variables. Furthermore, an evaluation of the R&D estimation results uncovered another error. Fabrizio and Tsolmon include time invariant variables that in theory should drop out when first-differencing the model, as showed in equation 3.6.

Table 5.2 reports the estimation of the R&D investment model without a constant term but still includes the time invariant terms, which gives us the opportunity to ex-amine the effect of including a constant term in the estimation of the R&D model as in table 5.1. The specifications clearly address the redundancy of adding a constant term when year indicator variables are included to the model. The specifications presented in columns two to five of table 5.2 are identical to the results in columns two to five of table 5.1, indicating that the year indicator variables capture the time trend in R&D expenditures. Comparing the results in the first and second column of table 5.1 and 5.2, we see that the addition of a constant does not has the same effect as including year indicator variables. Adding a constant decreases the coefficient of changes in industry output, while accounting for year effects, with or without addition of the constant,

Procyclicality of R&D investments : a critical study

Faculty of Economics and Business

Amsterdam School of Economics

Requirements thesis MSc in Econometrics.

Procyclicality of R&D Investments

a Critical Study

Gwendolyn Stuitje

Contents

Chapter 1

Introduction

Chapter 2

Literature Review

Chapter 3

Model

3.1

R&D Investment

3.2

Model Specification

Chapter 4

Data and Variables

4.1

Data

4.2

Construction of the Variables

4.3

Descriptive Statistics

Chapter 5

Replication of the R&D Model

5.1

Replication Results

5.2

Correcting the R&D Estimation