Faculty of Economics and Business
Amsterdam School of Economics
Requirements thesis MSc in Econometrics.
1. The thesis should have the nature of a scienti c 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
The procyclicality of R&D investment
a re-examination
Koos Gouweloos
ID: 10460578
MSc in Econometrics Specialisation: Free Track
Date of final version: December 28, 2016 Supervisor: Dr. J. C. M. Ophem
Second reader: Dr. K. J. van Garderen
Abstract
Empirical evidence shows that firms tend to invest in R&D during economic booms instead of during economic downfalls. This is also referred to as the procyclicality of R&D investment. Fabrizio and Tsolmon (2014) analyzed how patent effectiveness and the rate of obsolescence within an industry impact a firms R&D procyclicality. This research expands upon the research of Fabrizio and Tsolmon (2014) by taking their data and using more appropriate estimation methods. Specifically, these methods include fixed effects and random effects. Additionally we investigate whether there are difference in cyclicality between industries. This research finds a strong positive effect of patent protection levels on the cyclicality of R&D, something which goes against predictions. Furthermore this research finds no evidence of any effect of obsolescence on the cyclicality of R&D investment. We also find evidence that the impact of patent effectiveness and obsolescence on procyclicality can be different for different industry groups.
Statement of originality
This document is written by Koos Gouweloos 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 supervision of completion of the work, not for the contents.
Contents
1 Introduction 1 2 Literature Review 3 3 Model Specification 6 3.1 R&D investment . . . 6 3.1.1 First Difference . . . 73.1.2 Fixed Effects model . . . 7
3.1.3 Random Effects model . . . 8
3.2 Patent Output . . . 8
3.2.1 Poisson quasi-maximum likelihood estimator . . . 9
4 Data & Variables 10 4.1 The data . . . 10
4.2 Variables . . . 11
4.3 Descriptive Statistics . . . 14
5 Replication 15 5.1 Replicating R&D investment . . . 15
6 Results 19 6.1 Fixed Effects estimator . . . 19
6.2 Random Effects estimator . . . 20
6.2.1 Random vs Fixed effects . . . 21
6.2.2 Additional estimations . . . 21
7 Industry groups 27 7.1 Dummy estimation . . . 28
7.2 Estimation per group . . . 28
CONTENTS CONTENTS
Bibliography 32
Appendices 33
Chapter 1
Introduction
Traditional economic theory predicts R&D investment and innovation by firms to be anti-cyclical. Which means that spending on R&D is higher during economic downturns. The reasoning behind this is that firms should concentrate their innovative activities in recessions when opportunity costs are lower (Schumpeter, 1942). However empirical evidence finds that R&D investment and innovation seems to be pro-cyclical. Firms seem to innovate during economic booms and scale down on R&D during recessions (Fabrizio and Tsolmon (2014) , Barlevy (2007)). One explanation for this phenomenon has been that firms do not have the financial means to invest in R&D in times of recession. Empirical evidence by Aghion et al. (2012) shows that credit constraints can indeed play a role in recessionary periods. Another explanation for this procyclicality is that firms consciously time their innovation to coincide with economic booms. This timing is then done so that firms can reap the highest possible returns from their investments before imitators eat away at rents (Fabrizio and Tsolmon (2014), Shleifer (1986)).
One study of particular interest is An empirical examination of the procyclicality of R&D investment and innovation done by Fabrizio and Tsolmon (2014). In their paper they empirically tests a theory that R&D investment is procyclical because firms time their innovation to match times of economic upswings. They do this by looking at the rate of obsolescence and patent effectiveness within an industry and test how these influence the procyclicality of R&D investment and innovation. The rate of obsolescence within an industry is the rate at which products and processes get replaced by new ones. Firms operating in industries in which their innovation will be obsolete more quickly are expected to time their innovations during economic booms. Patent protection is a measure used to investigate the effect of imitation by rival firms on innovation and R&D investment. A higher level of patent protection would mean a lower risk of imitation, therefore making firms less procyclical. In the end they find evidence that firms in industries with a higher level of obsolescence are more procyclical. They did not find a
CHAPTER 1. INTRODUCTION significant effect for patent protection levels.
The models and techniques used by Fabrizio and Tsolmon to investigate this procycli-cality appear to be questionable. For instance they claim that their model is based on a similar fixed effects model done by Barlevy (2007), however upon inspection this does not seem to be the case. They specify a first difference model to estimate the effects of obsolescence and patent protection levels but they include time-invariant regressors to this model. Given that all the data used by Fabrizio and Tsolmon is publicly available on the data archive site of the Review of Economics and Statistics, this makes it possible to reproduce and expand upon the research done by Fabrizio and Tsolmon (2014) using different methods. Specifically Fixed effects and Random effects.
This thesis sets out expand upon the research done by Fabrizio and Tsolmon (2014). A fixed effect model and a random effects model using the data from the paper by Fabrizio and Tsolmon will be implemented. Additionally a closer look at differences between industries will be made. The results from these estimations will be compared to the results of Fabrizio and Tsolmon. It will be of interest to see whether the same conclusions still hold and hopefully additional interesting results can be found.
The remainder of this thesis is structured as follows. Chapter 2 gives a literature review of the cyclicality of R&D investment. The next chapter 3 describes the methods and methodology used in this research, followed by chapter 4 which describes the data. Chapter 5 describes how well the original results found by Fabrizio and Tsolmon (2014) can be replicated. Chapter 6 presents the results of the random and fixed effects estima-tion. In Chapter 7 we look at differences in cyclicality between industry groups. Finally chapter 8 concludes.
Chapter 2
Literature Review
This thesis examines how different levels of patent protection and obsolescence within an industry, affect a firms procyclicality when it comes to R&D investment. The relevant literature on this subject will be discussed in the following section.
One of the most important researcher when it comes to the analysis of business fluc-tuations and innovation is Schumpeter. He introduced the theory of creative destruction (Schumpeter, 1942). This theory describes an incessant product and process innovation system by which old, outdated products get replaced new ones. This restructuring pro-cess would increase during repro-cessions and indicate a potential counter-cyclical behavior (Caballero, 2010). Schumpeter reasons that recessions even play a useful role in growth by forcing firms to restructure and remove inefficient processes and products.
However, most evidence in the theoretical and empirical literature points towards that firms act pro-cyclical. R&D investment and innovation are high in times of economic prosperity and decrease in times of recession. Barlevy (2007) estimates a fixed effects model and finds that R&D investment is pro-cyclical. According to Barlevy even though it is socially optimal if entrepreneurs concentrate their innovation in recessions, external forces makes them short-sighted. Risk of imitation by competitors makes them focus on short-term profits and delay investment in R&D until during an economic boom. An important assumption Barlevy makes in his research is that R&D investment and innovation happen at the same time, they are contemporaneous. Some other researchers see R&D investment and innovation as two separate strategic decisions.
Shleifer (1986) describes a theoretical model in which inventions arrive in different sectors at different times exogenously. The firms then decide at which time they im-plement these inventions. Once an innovation is imim-plemented other firms can imitate the invention and eat away the profits. Schleifer reasons that this threat of imitation causes firms to wait with implementation until periods when its profits are going to be the highest, which is during an economic boom. When firms all have similar expectations
CHAPTER 2. LITERATURE REVIEW about when this boom will occur they can time their innovations to make it a reality. Clustering the implementation of innovations will cause a surge in labor demand which in turn leads to high demand. This self reinforcing cycle supports the procyclicality of R&D investment.
Building upon the insights of Schumpeter and Shleifer is a paper written by Francois and Lloyd-Ellis (2003). They introduce a theory that unlike Shleifer has endogenous cyclicality. They find that firms will do the R&D during recessions and then store the inventions until there is an economic boom. So that R&D spending would actually be counter-cyclical.
One of the most recent contributions to the literature is done by Hud and Rammer (2015). They investigated whether firms that increased their innovation budget after the global economic crisis of 2008 were able to generate higher innovative rewards in the following period. But they found mixed results. At the industry level they find no significant counter-cyclical effects, indicating that it is not a good strategy for a firm to increase R&D budget when their primary sales market is going down. However at the GDP level they do find a positive effect of expansive innovation budgeting during the economic crisis.
Fabrizio and Tsolmon (2014) build upon the research done by Barlevy (2007), but question the assumption R&D and innovation being contemporaneous. They say that the timing of innovation and R&D investment by firms could be two separate strategic decisions and they therefore model them separately. Firms first decide to do research and to find problems and find solutions. This leads to inventions which after further develop-ment and commercialization are turned into innovations. Besides that general demand influences these decisions, these decisions are also influenced by the rate of obsolescence and the level of patent protection within an industry. In the next part only the R&D investment part of Fabrizio and Tsolmon will be discussed.
Imitation by competitors is of influence when deciding to invest in R&D. Firms invest in order to reap profits and rewards over this investment. To protect these investments from imitation by competitors, firms try to patent their inventions. How well a firm can protect it inventions by patenting differs across industries and it are these patent protection levels that can possibly influence a firm’s decision when to invest in R&D. According to Fabrizio and Tsolmon (2014) there are two scenario’s. If investment in R&D and innovation are contemporaneous than firms in industries with a lower level of patent protection would be inclined to shift R&D to times when demand is high. Fabrizio and Tsolmon give the following corresponding hypothesis.
• R&D investment will be more sensitive to changes in demand for firms in industries with weaker patent protection, relative to firms in other industries all else equal.
CHAPTER 2. LITERATURE REVIEW A second scenario is that R&D and innovation are not contemporaneous, meaning firms can first do R&D and then wait with introducing inventions from this research when they see fit. In that case Fabrizio and Tsolmon argue that patent protection levels would not have an effect on the procyclicality of R&D investment.
Obsolescence is the rate at which new products become obsolete and are replaced with new products. A new invention will then loose value regardless whether it is introduced to the market or not. This rate can also affect the timing of R&D investment. Fabrizio and Tsolmon reason that a higher rate of obsolescence will lead firms to increase research and development during booms so that any inventions created can be introduced swiftly when demand is high. They state the following hypothesis:
• R&D investment will be more sensitive to changes in demand for firms in industries with faster obsolescence of products, relative to firms in other industries, all else equal.
To test these hypothesis Fabrizio and Tsolmon use a large panel data set of US manufac-turing firms. The sample period spans from 1975 till 2002. They estimate the following first difference equation:
∆RDkt= β0+ β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+ β4Obs + β5∆Xit× Obsi
+β6P atEf f + β7∆Xit× P atEf fi+
X
τt+ kt
(2.1) where ∆RDktis the growth in firm-level R&D expenditures for a firm k in year t, ∆Xit is
industry growth of industry i in year t, Obsi and P atef fi are two variables that measure
the rate of obsolescence and the level of patent protection within an industry i. The variables ∆Mkt and ∆Mk,t−1 represent a set of control variables and their lag. P τt
are year indicator variables included to capture time effects. Time effects are added to control for omitted variables that are common between all firms but that vary over time. In chapter 4 a more detailed description for these variables will be provided.
In the above equation the interaction terms of Output with the variables Obsi and
P atEf fi are to test the two previously mentioned hypothesizes. A positive coefficient for
β5 would imply that firms in industries with a faster rate of obsolescence behave more
procyclical. A negative coefficient for β6 confirms that firms in industries with higher
patent protection levels are less procyclical.
After estimating their model Fabrizio and Tsolmon (2014) do not find a significant value for β6 rejecting their first hypothesis. This however could imply that R&D and
innovation are not contemporaneous. Fabrizio and Tsolmon find a negative coefficient for β4 confirming their second hypothesis.
Chapter 3
Model Specification
This section gives a summary of the models used in this research. Specifically the models used to estimate the impact of changes in industry output on investments in R&D and the model used to estimate a patent equation.
3.1
R&D investment
The data used in this research are repeated observations on the same cross section, observed for multiple time periods also known as panel data. Typically analyzed using linear panel models. Suppose the panel data consists of observations on N cross-sectional units (firms) and T time periods. A very general model for this is:
yit= α + ci+ x0itβ + z 0
iγ + it (3.1)
where yit is the dependent variable, xit is a matrix of explanatory time-variant
vari-ables of interest and zi are observable time invariant variables. The ci’s are individual
specific effects and they model the impact of time-invariant heterogeneity between in-dividuals that can not be observed. When estimating β, it is important to specify the nature of ci. One variant of Model 3.1, the so called Fixed Effects model, treats ci as
an unobserved random variable that is possibly correlated with the explanatory variables in xit (Cameron and Trivedi, 2005). Another case is when ci and xit are unrelated and
independently distributed, this is called Random Effects. In this instance the ci’s are
treated as part of the error term.
The distinction between Fixed Effects and Random Effects is an important one. When fixed effects are present and the ci’s are correlated with the regressors xit, some estimators
like ordinary least squares (OLS) are inconsistent. For observational data, the omitted and included variables are often correlated with one another. Therefore the general opinion is that RE effect models are unrealistic.
3.1. R&D INVESTMENT CHAPTER 3. MODEL SPECIFICATION Translating the general Equation 3.1 to the model estimated in this research:
RDkt= α + ci+ β1Mkt+ β2Mkt−1+ β3Xit+ β4Xit× Obsi
+β5Xit× P atEf fi+ γ1Obsi+ γ2P atEf fi
+Xτt+ kt
(3.2) The next section describes the first difference estimator and the model implemented by Fabrizio and Tsolmon (2014), followed by the Fixed Effects estimator and the Random Effects estimator.
3.1.1
First Difference
In the presence of fixed effects, one way to get consistent estimates is first differencing Equation 3.1. Lagging 3.1 one period gives yi,t−1 = α + ci + x0it,t−1β + zi0γ + i,t−1, when
subtracting this from the original Equation 3.1, the first difference model is:
yit− yi,t−1 = (xit− x0i,t−1)β + (it− i,t−1) (3.3)
The ci’s have dropped out and the OLS estimator of equation 3.3 will be consistent. Using
first differencing means that the constant α and the parameters γ for the time invariant regressors zi will not be identified. Now let us look a second time at the first difference
model by Fabrizio and Tsolmon (2014) as specified in Equation 2.1
One problem with this model specification of Fabrizio and Tsolmon is the inclusion of β0, a constant, which should have dropped out after first differencing. Secondly the
inclusion of the variables P atEf fi and Obsi is curious, since these are time invariant.
Why they decided to add these time invariant regressors to their first difference model is not clear. One possible explanation is that they were very keen on identifying β5 and β7.
However for testing their hypotheses one only needs to look at the interaction terms of Patent Effectiveness and Obsolescence with Output. A more appropriate first difference model would have been:
∆RDkt= β1∆Mkt+ β2∆Mkt−1+ β3∆Xit+ β4∆Xit× Obsi
+β5∆Xit× P atEf fi+
X
τt+ kt
3.1.2
Fixed Effects model
A second estimation method when dealing with fixed effects is the within estimator. It exploits the special features of panel data. It uses the cross-sectional information of the differences between firms and uses the within-firm information of differences over time. Looking again at the general model
yit= α + ci+ x0itβ + z 0 iγ + it
3.2. PATENT OUTPUT CHAPTER 3. MODEL SPECIFICATION Averaging over time:
yi = α + ci+ x0iβ + z 0 iγ + i (3.4) where yi = PT t=1yit, xi = PT t=1xit, and i = PT
t=1it. Subtracting both these
expres-sions leads to:
yit− yi = (xit− xi)0β + it− i (3.5)
As can be seen the ci’s drop out. One drawback again of this method again is that
estimation of time-invariant variables zi will not be possible. For instance variables
like Patent Protection and Obsolescence. Furthermore even for time varying variables estimates can be imprecise if a lot of variation in these variables are cross-sectional rather than over time (Cameron and Trivedi, 2005). The within estimator is then the OLS estimator of equation 3.5 and will give consistent estimates when fixed effects are present. However, correction for possible heteroscedasticity and serial correlation in the errors is needed. If not, reported standard errors can be smaller than they actually are. To circumvent this problem the fixed effects model is estimated in Stata using cluster-robust estimation.
3.1.3
Random Effects model
Finally this research also estimates a random effects model. Rewriting Equation 3.1 as: yit = α + ci+ x0itβ + z
0
iγ + vit (3.6)
where vit = ci+ it. This model assumes that ci and it are i.d.d and so treats ci as part
of the error term. If these assumptions are correct than OLS will be consistent. Again to avoid possible inflated standard errors cluster-robust estimation is done. An advantage of random effects is that time invariant regressors can now be identified. However it is important to keep in mind that most economists however when facing observational data choose not to rely on Random Effects because of its strong assumptions (Cameron and Trivedi, 2005).
3.2
Patent Output
Although this research’s main focus is on Fabrizio and Tsolmon’ R&D investment model, it is still interesting to replicate their results from modeling the number of patent applica-tions. This serves as an extra check if the results reported in their paper are reproducible. Therefore this section provides a brief description of the Poisson quasi-maximum likeli-hood estimator.
3.2. PATENT OUTPUT CHAPTER 3. MODEL SPECIFICATION
3.2.1
Poisson quasi-maximum likelihood estimator
Firm-level patent output is modeled in order to find out how firms in industries with varying rates of obsolescence and levels of patent protection time their patent applications. Since the dependent variable is a count, a Poisson maximum likelihood estimator is chosen. This model is then estimated using a fixed effects estimator. Fabrizio and Tsolmon estimated the following patent equation:
E[Pkt|Xit, Zkt−1] = exp(β1RDkt−1+ β2Xit+ β3Mkt−1+ τt+ µk)
which they based on a model used by Hall et al. (1986) and it estimates the relationship between patented inventions in one year (Pkt) and R&D investment in the previous year
(RDkt−1). The coefficient for industry output Xit measures whether a firm applies for
more or less patents as industry output changes. Furthermore one-period lagged firm-level controls (Mkt−1) are added to the model. µk and τk are firm and year effects.
Then additionally a model is estimated including two interactions terms with industry output. Namely the rate of obsolescence (Obsi) and patent effectiveness (P atef fi). This
is done in order to specifically measure how these two conditions influence the relationship between industry output and number of patents a firm generates. The final model is: E[Pkt|Xit, Zkt−1] = exp(β1RDkt−1+β2Xit+β3Mkt−1+β4Xit×Obsi+β5Xit×P atEf fi+τt+µk)
It is interesting to see that Fabrizio and Tsolmon did leave out any time constant estimators in this estimation. In an industry with a high rate of obsolescence it is expected that firms have less incentive to delay innovation and therefore a negative coefficient β4
for the interaction of Obsi and Xit is expected. Furthermore it is expected that firms in
industries with low levels of patent protection are more sensitive to changes in industry output. Therefore the coefficient β5 is expected to be negative as well.
Chapter 4
Data & Variables
In the following section the data and variables used in this research are described. First the different data sources are described followed by an explanation how, from this data the estimation samples used are created. After that, there is a description of each variable used in the analysis and how to obtain it from the data. Finally some descriptive statistics are provided.
4.1
The data
The data used in this thesis come from the The Review of Economics and Statistics’ online database. It was made publicly available by Fabrizio and Tsolmon. In this section the data and variables they used in their research will be described. The data comes from four different sources.
Firstly use is made of the NBER Patent Data and secondly industry-level annual demand data from the NBER Manufacturing and Productivity database. The National Bureau of Economic Research is an American private non-profit economic research orga-nization.
Furthermore data is used from the Carnegie Mellon survey (CMS) of R&D managers. This is a privately conducted survey from 1994 which collected information from man-ufacturing firms’ R&D managers on the determinants of R&D activity. The data was collected on topics as firms’ use and perceptions of the effectiveness of different mech-anisms for protecting intellectual property in order to appropriate the returns to their innovations, sources of new technical information, interactions with competitors, and the speed at which new products or processes are imitated. This survey was sent out to 3240 R&D labs in the United States of which 1478 responded. The CMS is used to create the industry level measures Obs and Pateff. The responding firms were both public firms and private firms.
4.2. VARIABLES CHAPTER 4. DATA & VARIABLES The three data sources were combined with data from Compustat. This is a market database published by Standard and Poors, it provides company information going back as far as 50 years.
From these four data sources two samples were created. One sample is used to model firm patent applications and one sample is used to model R&D investment. The patent sample was created by starting with all firms listed in Compustat from 1975 to 2002. Those that where matchable were combined with the NBER patent data using the firm identifier (gvkey). Observations that had no number of patents and observations that had missing values for control variables were dropped. Furthermore cases with only one firm-year observations per firm were dropped as well. Matching the remaining firms with the CMS survey trimmed down the data set even more. The final estimation sample is an unbalanced data set that contains 48,477 observations, which translates into 711,570 patents by 4,029 firms in 101 different manufacturing industries from 1975 to 2002.
The data set for the R&D investment model is created similarly. Again starting with all firms listed in Compustat 1975 and 2002. These firms are matched by industry to the NBER Manufacturing and Productivity database. Industries are coded by three digit SIC codes and firms were matched by industry to industry-level annual output data. Firms for which there are no values for R&D expenditures and for which control variables were missing, were dropped. This sample is then matched to CMS survey, further reducing the data set. This second sample contains then 71,264 observations from 7,731 different firms in 102 different manufacturing industries.
4.2
Variables
R&D growth. The log number of firm-level R&D expenditures in a given year, pro-vided by Compustat. As mentioned previously Compustat only covers publicly traded firms, and these firms tend to be large. Barlevy (2007) compares the Compustat R&D expenditures to expenditure data from the National Science Foundation (NSF) which includes a more representative sample of firms. He finds very few differences in R&D growth between the two data sets, which gives confidence that the Compustat is actually quite representative.
Industry output growth. The natural log of output for a given year in a given industry. This variable is constructed again using the NBER Manufacturing and Pro-ductivity database. Following Barlevy (2007), for each different industry output was calculated by summing annual value-added and material costs. This was then divided by each industries shipments deflator. This output data only covers manufacturing indus-tries although this means that certain R&D intensive indusindus-tries are left out (for example
4.2. VARIABLES CHAPTER 4. DATA & VARIABLES computer programming and communications), however in general R&D is very heavily concentrated in manufacturing.
Patent Effectiveness. This is a measure for the effectiveness of patent protection within an industry and is constructed using the CMS survey. This survey asked the participants for the percentage of product and process innovations for which patenting was effective in protecting their firms competitive advantage associated with those in-novations. Respondents had to respond for product and process innovations separately and could choose between the following options: under 10%, 10-40%, 41-60%, 61-90% and above 90%. To get a single measure out of these answers, Fabrizio and Tsolmon computed a weighted average of the product and process scores. Using R&D percentage spent on product and process innovation to weigh? These scores are then averaged to the industry-level to get the industry level variable for patent effectiveness. The higher the value of this variable the higher the (perceived) patent protection effectiveness in the industry.
Rate of obsolescence. To find a measure for obsolescence Fabrizio and Tsolmon one again turn to the CMS survey, which asked participants to indicate the speed with which products and with which processes are introduced in their industry. To do so they could choose between five options Very Slowly, Slowly, Moderately, Rapidly or Very Rapidly. Each response is respectively assigned a number from 1 to 5 and again a weighted aver-age is calculated using percentaver-age of R&D spending on product and process innovation. These scores are averaged at the industry level and this is the variable used to measure obsolescence within an industry.
Fabrizio and Tsolmon mention that within an industry responses to this survey vary. Therefore it is important to check how big this variation within industries is in order to make sure that averaged industry score is representative for the majority of firms within that industry. In their paper they provided 2 histogram of responses for two different industries. These histograms can be seen in figure 4.1. Where 4.1a shows the responses for an industry with a slow rate of obsolescence and 4.1b shows an industry with a fast rate of obsolescence. It can be seen in the figures that the majority of the responses are in line with the industry classification. Do note that the R&D sample involves 118 different manufacturing industries and Fabrizio and Tsolmon only made aggregate data available. It would have been interesting to check additional industries responses distributions besides these two.
External financing. It is important to control for differences between industries in dependency on external financing. Having good or bad access to external funds can affect a firms R&D investment levels. To construct a measure of external financing, Fabrizio and Tsolmon (2014) look at a firms total external funds needed to finance its
4.2. VARIABLES CHAPTER 4. DATA & VARIABLES
(a) Slow obsolescence industry (b) Fast obsolescence industry
Figure 4.1: Distribution of survey responses for two industries
investments over our sample years. Using the Compustat data. This sum is then divided by the firms total capital expenditures over the sample years . This ratio is represents a firms dependence on external financing. Now to get an industry level variable of external financing the median value of all the firms within a certain industry is used.
Firm level controls. For the R&D investment model the firm level controls are taken from a previous study done by Barlevy (2007), in his study he controlled for a firms balance sheet positions. Reasoning that when a firm’s earnings fall in recessions it will be harder to finance R&D. The variables included are taken from Compustat. They are all natural logs and depending on which model is estimated we take the first difference of: one period lagged and contemporaneous cash flow before R&D expenditures (Cash Flow), contemporaneous and one-person-lagged measures of total assets (Total Assets), total liabilities (Total Liabilities), long term and short-term debt (LT Debt and ST Debt), and capital stock (PPE).
For the Patent model a different set of controls is used. The size of the company is controlled for using the log of total number of firm employees (ln EMP). Additionally there is a control for asset intensity which is represented by the one-period-lagged natural log of firm plant, property and equipment value (ln PPE). Furthermore the effect of annual sales on patenting is controlled for by including the one-period-lagged natural log of sales(ln SALES).
Time dummies are included to control for omitted variables that are common to all firms but that change over time. One could think of broad technological changes or changes in state policy that affect all companies. Time dummies are included for the years 1975 to 2002.
4.3. DESCRIPTIVE STATISTICS CHAPTER 4. DATA & VARIABLES
4.3
Descriptive Statistics
In table 4.1 a summary of the variables used in this research is provided. The R&D investment model equation was estimated using a sample of 78,500 observations for 8,213 firms covering the years 1975 till 2002. The dependent variable of this research, the nat-ural log of R&D output ranges from 0 to 13.76 with a median value of 3.81. Furthermore let us look at the two most important explanatory variables in this research, Obsolescence and Patent Effectiveness.
The rate of obsolescence ranges from 1.56 to 4 with an average of 2.85, which translates to about a moderate speed. Industries with slow rates of obsolescence are for instance Railroad Equipment (SIC 295) with a score of 1.75 and Asphalt Paving and Roofing Materials which scored 1.56. Very fast industries are Blankbooks, Looseleaf Binders, and Bookbinding (SIC 278) and Computer and Office Equipment (SIC 357) with scores of respectively 4 and 3.55.
Patent Effectiveness ranges from 0.05 to 0.82. Industries which have a high level of patent protection are Transportation Equipment (SIC 379) and a score of 82% and Agricultural Chemicals (SIC 287), with a score of 75%. Low levels of patent protection were found in industries like: Publishing and Printing (SIC 272), Cigarettes (SIC 211), and Textile Products (SIC 239). Each of which had a score of 5%.
Table 4.1: Descriptive statistics
N N (firms) Mean Median SD Minimum Maximum R&D 78500 8213 3.56 3.81 3.45 0 13.76 Cash Flow 78500 8213 7.77 7.26 1.40 -1.66 15.14 Total Assets 78500 8213 8.52 8.68 3.03 0 17.39 Total Liabilities 78500 8213 7.73 7.77 3.02 0 17.37 LT Debt 78500 8213 5.68 6.14 3.81 0 16.37 ST Debt 78500 8213 3.71 3.88 2.90 0 14.80 Capital Stock 78500 8213 7.04 7.1 3.18 0 16.08 Output 78500 8213 8.93 8.75 1.06 5.79 12.37 Obsolescence 78500 8213 2.85 2.87 0.35 1.56 4 Patent Effectiveness 78500 8213 0.44 0.45 0.15 0.05 0.82 External Financing 78500 8213 0.37 -0.14 1.42 -2.93 4.76 Year 1975 2002
Chapter 5
Replication
Fabrizio and Tsolmon made all the data used in their paper publicly available, along with a Stata ”do file” which documents all the commands used to find their final results. These files can be found in the online database of The Review of Economics and Statistics. This chapter checks whether we can actually reproduce the results found before we present the result of the Random and Fixed effects models. A check which can be done first is to re-estimate the Innovation Model by Fabrizio and Tsolmon (2014). In their paper they modeled:
• Number of patents a firm applied for in a year t depending on industry output • Firm-level expenditure on R&D as function of changes in industry output.
Re-estimating the patent equation was not a hard task. A final estimation sample was provided by the authors and it was just a matter of rerunning the equations provided in the Stata do-file. In the appendix in Table A.1 the results can be seen plus a copy of Table 4 from the original paper. As can be seen the results are exactly the same. This serves as a validation that the authors provided the correct data to the online database, however this needs to be checked for the R&D equation also which is of main interest of this research.
5.1
Replicating R&D investment
Again using the data and code provided by Fabrizio and Tsolmon (2014), we run the analyses for the R&D Model. Trying to recreate the results reported in Table 5 of their paper and see if the same conclusions can be made. The results of this can be seen in Table 5.1. In the appendix in Table A.2, one can see the original results of Fabrizio and Tsolmon.
5.1. REPLICATING R&D INVESTMENT CHAPTER 5. REPLICATION We run a pooled regression with robust standard errors clustered by firm. This esti-mation reports 71,264 observation used in the regression, the same number as reported by Fabrizio and Tsolmon. Overall the result are similar but not exactly the same. All coefficients have the same sign and significance but they differ in magnitude (however for many control variables the results are nearly identical), this difference grows as we look through the table from left to right and are most profound when looking at the variable delta Output. How these differences arise is unclear. According to explanatory files by Fabrizio and Tsolmon in the database of The Review of Economics and Statistics, the data provided should be exactly the same as the data used for their own estimations. The number of observations to produce both tables is identical. So there is no easy way to check how these small differences arise.
Looking at the results in table 5.1, similar conclusions can be drawn as in Fabrizio and Tsolmon. The first column presents the effect of Output growth and other control variables on R&D investment, but year indicators are not included. The coefficient of the Output variable is positive and significant with a magnitude of 0.204 versus 0.211 in Fabrizio and Tsolmon (2014). A difference of 0.007. A positive coefficient for Output supports the procyclicality of R&D investment and is in itself. Firms seem to invest in R&D when Output in their industry is high. In the second column year indicators are included in the estimation, this however does not change the coefficients much (0.272 vs 0.280). Fabrizio and Tsolmon reason that ”The estimated procyclicality of R&D was not diminished by this set of controls, and the fact that the R2 is unchanged by inclusion
suggests that the common year trends do not have significant explanatory power across industry”. This holds in the results of table 5.1 as well.
The third column shows the results of testing the first hypothesis mentioned of Fabrizio and Tsolmon (2014), which goes: ”R&D investment will be more sensitive to changes in demand for firms in industries with weaker patent protection, relative to firms in other industries, all else equal”. This is done by now including variables for Patent Effectiveness and an interaction term Output × Patent Effectiveness. The coefficient for Patent Effectiveness is positive and significant, providing evidence that industries with stronger patent protection have higher R&D spending. With a magnitude of 0.172 it is almost identical to the estimate of Fabrizio and Tsolmon of 0.175. The interaction of Output with the indicator of patent protection (Output × Patent Effectiveness) is not significant. Therefore the theory that R&D investment will be more sensitive to changes in demand in industries with weaker patent protection is unfounded.
In the fourth column we run the model including variables Obsolescence and the in-teraction Output × Obsolescence. This is done to test hypothesis 4, R&D investment will be more sensitive to changes in demand for industries with faster obsolescence of
prod-5.1. REPLICATING R&D INVESTMENT CHAPTER 5. REPLICATION ucts, relative to firms in other industries, all else equal. Again the coefficients are almost identical to the ones of Fabrizio and Tsolmon. Both these coefficients are positive and significant and according to Fabrizio and Tsolmon (2014) this supports their predictions that R&D investments were more procyclical in these industries.
The fifth column represents their full model, including both sets of interactions. We find that Output × Patent Effectiveness is −0, 054 but insignificant and Output × Obso-lescence is 0.341 and significant. Compared to respectively −0.109 and 0.309 in Fabrizio and Tsolmon (2014). Here the differences are a bit bigger. The coefficient for Obsoles-cence is about twice as large as the one reported in Fabrizio and Tsolmon (2014).
The sixth column adds an industry level variable to the full model which controls for reliance on external financing. Interacted with changes in demand (Output × External Financing) in order to account for industries that might be financially constrained to a greater extend during recessionary periods. The coefficient for the interaction term is not significant in both Table 5.1 and Table 5 of Fabrizio and Tsolmon (2014). In the final column year fixed effects are excluded.
All in all when looking at the replication of Fabrizio and Tsolmons results it can be seen that they are, for the most part, very close. Especially when looking at the significant coefficients the differences are not that large. For instance the difference between the coefficients of Output growth in the first column is only 0.007 (0.211 versus 0.204) while in the fifth column this difference is 0.120 (-0.657 versus -0.777). The coefficients for the interaction terms are even closer, with the biggest difference between coefficients being 0.032. This gives confidence that the data provided is mostly the same and that we can continue the analysis.
5.1. REPLICATING R&D INVESTMENT CHAPTER 5. REPLICATION Table 5.1: OLS estimates in R&D Expenditures
(1) (2) (3) (4) (5) (6) (7) ∆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) Patent 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) ∆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) ∆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) ∆Total Assets 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) ∆Total Assets 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) ∆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) ∆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) ∆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) ∆LT Debt lag 0.000 0.000 0.000 0.000 0.000 0.002 0.000 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) ∆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) ∆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) ∆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) ∆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) C 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) 71,264 71,264 71,264 71,264 71,264 71,264 71,264
R2 0.22 0.23 0.23 0.23 0.23 0.23
Chapter 6
Results
This section presents the results of the Fixed and Random effects analysis.
6.1
Fixed Effects estimator
As previously explained in Chapter 3, the model estimated by Fabrizio and Tsolmon has a questionable specification. They only take the first difference of a limited set of variables and then include a constant and add variables that do not change over time. Table 6.1 shows the results of estimating their model without first differencing and instead using fixed effects. One of the first things to notice is that the variables that do not change over time, Patent Effectiveness, Obsolescence and External Financing immediately drop out.
Taking a closer look at the results. The first two columns are similar to the ones reported in Fabrizio and Tsolmon. With a positive and significant coefficient for Output at 0.281 compared to 0.211 in column 1. Also mostly the same set of control variables are significant and have the same sign.
In the third column which reports the result of the first hypothesis the first difference with Fabrizio and Tsolmon can be seen. First of all the variable Patent Effectiveness is dropped out because it does not change over time. Secondly the interaction Output × Patent Effectiveness is positive and significant. This is a different result and it goes against what Fabrizio and Tsolmon found which was a slightly positive but insignificant coefficient. Furthermore it is the opposite of what they expected to find. Firms in industries with lower patent protection levels should be more sensitive to changes in demand. Moreover the effect seems to be quite large with a magnitude of 1.518.
The fourth column reports the results of testing hypothesis 2 which predicted that R&D investment is more sensitive to changes in demand for industries with faster rates of obsolescence. Using fixed effects no evidence for this can be found. The coefficient is
6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS not significant. Which is curious since this is the main result of Fabrizio and Tsolmon (2014). As stated in the previous section they found no evidence for patent protection levels interacting with Output to influence R&D investment, but they did find a positive effect for Obsolescence. The fifth column which shows the results of the full model gives the same conclusions.
In the sixth and seventh column results are reported of the full model including controls and interactions for External Financing. It can be seen that there is a strong and positive effect which is something not found using the difference model of Fabrizio and Tsolmon (2014). Using fixed effects it seems that there is evidence that greater demand leads to greater availability of external funds which leads to bigger R&D investment.
Overall what can be seen estimating the model using fixed effects is that the two hypotheses made by Fabrizio and Tsolmon in their paper regarding R&D investment do not hold. Although the interactions of Output with Patent Effectiveness are significant it has not the right sign. The interaction of Obsolescence with Output is insignificant. Rejecting their hypotheses. However the coefficient for Output remains positive and significant, so their conclusions that R&D investment is procyclical still holds.
6.2
Random Effects estimator
Now presenting the result of using a random effects estimator. Using the random effects estimator requires that the individual specific effects αi are independently distributed
from the regressors. Which is a very strong assumption to make. However the random effects estimator does allow for the estimation of time invariant variables, unlike the fixed effects estimator. Something which must have been important for Fabrizio and Tsolmon (2014) since they decided to wrongfully include the time-invariant regressors Patent Effectiveness, Obsolescence and External Financing in their first difference model. In table 6.2 the result of running the model using random effects can be seen. The tables layout is the same to the previous ones with each column providing the results of a different estimation. In the first and second columns, which shows the results of estimating R&D as a function of only Output and a set of control variables, it can be seen that R&D is procyclical. With twice a significant and positive coefficient of 0.298 and 0.242 respectively. The coefficients of Output are of the same order and significance as in the FE estimator and Fabrizios first difference estimator. Also again the same control variables are significant.
In the third column, which reports the results of testing the first hypothesis it can be seen that the interaction of Output and Patent Effectiveness is significant and positive. Much the same as in the FE model. The coefficient of just Patent Effectiveness is
signif-6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS icant as well and very strong. Curiously it has the opposite sign of the interaction term. Which would mean stronger patent effectiveness lead to less investment in R&D overall. The fourth column once again reports the results of testing the second hypothesis. The interaction term Output × Obsolescence is positive and significant. In line with the original results of Fabrizio and Tsolmon. The coefficient for Obsolescence however is not significant.
The fifth column, which reports estimating R&D investment with both interactions of Patent Effectiveness and Obsolescence included, shows that both coefficients are positive and significant. A result different from both Fabrizio and Tsolmon and the fixed effects model. In columns six and seven none of the variables of interest, Output and its interac-tions with Patent Effectiveness and Obsolescence are significant. Only the interaction of Output with External Finance is positive and significant similarly as in the fixed effects model. The coefficient of External Finance is −1.341 and also significant.
6.2.1
Random vs Fixed effects
Using both random and fixed effects a significant positive coefficient is found for the interaction of Output and Patent Effectiveness. This indicates that patent protection levels can increase a firms procyclicality. An interesting result. For the interaction of Output and Obsolescence random effects does find a positive effect but fixed effects does not.
Random effects relies on the assumption that fixed effects are not present, if they are than the estimates of the random effects estimation are biased. In order to decide between fixed or random effects a Hausman test is run. This test indicates that fixed effects are present (prob > chi2 = 0.0000). So the results for the random effects estimator are possibly biased.
6.2.2
Additional estimations
Overall using either random effects or fixed effects creates different results from Fabrizio and Tsolmon (2014). There is a significant and positive effect of patent protection levels on cyclicality found both using fixed as well as random effects. No relationship is found for obsolescence using fixed effects. Random effects finds some evidence for a positive effect of obsolescence but when External Financing is included, obsolescence becomes insignificant. The results for the set of controls are mostly the same no matter which model is used.
6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS Subsample around survey date
One limitation of this research is that the industry conditions Obsolescence and Patent Effectiveness are treated as constant over the sample period. As mentioned these variables were created from a single survey conducted in 1994. However it might not be realistic to assume that perceived industry conditions in 1994 are the same as in 1975 or 2002. This could also explain why no significant effects for Obsolescence was found using the fixed effects model. Therefore the fixed effects model is re-estimated using only years closer to the date of the survey. Included are three years before and three years after the CMS survey, so a period ranging from 1991 till 1997. The results of this estimation can be seen in Table 6.3.
Again a positive and strongly significant effect is found for Patent Effectiveness. In column 2 a positive and significant effect is found for Obsolescence something not found estimating the full sample. However estimating the full models in columns 3 till 5 the coefficient for Obsolescence becomes insignificant again. It seems taking a sample closer to the survey date does not change the conclusions made in the previous section.
Including employees and sales
For their patent equation Fabrizio and Tsolmon (2014) use a different set of control variables than in the R&D investment model. Two of these controls are the natural log of total firm employees and the natural log of sales. One would expect that bigger firms with more employees would spend more on R&D than smaller firms. Similarly, higher sales could mean higher R&D spending. For this reason we re-estimate Equation 3.2 but this time including these two additional variables. The results of which can be seen in Table 6.4.
It can be seen that Employees and Sales are both positive and significant. Especially the number of employees seems to have a strong effect on R&D investment, having a magnitude ranging from 0.355 to 0.371 depending which interactions are added in the model. However the coefficients for our variables of interest, the interactions with Output, do not change much by including these variables. Our results from Section 6.1 still hold. Perhaps this is the reason that Fabrizio and Tsolmon (2014) did not include them in their R&D estimation.
6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS Table 6.1: Fixed Effects estimates in R&D Expenditures
(1) (2) (3) (4) (5) (6) (7) Output 0.281* 0.236* −0.325* −0.046 0.009 0.300 0.306
(0.029) (0.035) (0.100) (0.247) (0.249) (0.254) (0.243) Output × Patent Effectiveness 1.518* 1.678* 0.939* 0.985* (0.255) (0.278) (0.342) (0.339) Patent Effectiveness
Output × Obsolescence 0.089 −0.124 −0.149 −0.154 (0.076) (0.084) (0.082) (0.079) Obsolescence
Output × External Financing 0.174* 0.181* (0.043) (0.043) External Financing
Cash Flow 0.012 0.011 0.009 0.011 0.007 0.006 0.005 (0.018) (0.018) (0.018) (0.018) (0.018) (0.018) (0.018) Cash Flow lag 0.095* 0.087* 0.089* 0.088* 0.088* 0.089* 0.092* (0.018) (0.017) (0.017) (0.017) (0.017) (0.017) (0.018) Total Assets 0.250* 0.253* 0.256* 0.254* 0.256* 0.255* 0.255* (0.020) (0.020) (0.020) (0.020) (0.020) (0.020) (0.020) Total Assets lag 0.126* 0.130* 0.129* 0.130* 0.129* 0.128* 0.125* (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) Total Liabilities 0.087* 0.081* 0.078* 0.081* 0.078* 0.077* 0.079* (0.019) (0.020) (0.020) (0.020) (0.020) (0.020) (0.019) Total Liabilities lag −0.115* −0.120* −0.120* −0.120* −0.120* −0.120* −0.118* (0.014) (0.014) (0.014) (0.014) (0.014) (0.014) (0.014) LT Debt −0.033* −0.031* −0.031* −0.031* −0.031* −0.030* −0.031* (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) LT Debt lag 0.003 0.003 0.004 0.003 0.004 0.004 0.004 (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) ST Debt −0.001 −0.002 −0.002 −0.002 −0.002 −0.002 −0.001 (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) ST Debt lag 0.004 0.003 0.003 0.003 0.003 0.003 0.003 (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) Capital Stock 0.121* 0.122* 0.122* 0.123* 0.122* 0.123* 0.123* (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) Capital Stock lag 0.009 0.008 0.007 0.008 0.007 0.006 0.007
(0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) Year fixed effects No Yes Yes Yes Yes Yes No N (observations) 78,500 78,500 78,500 78,500 78,500 78,500 78,500 R2 0.24 0.24 0.20 0.25 0.18 0.16 0.15
6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS Table 6.2: Random Effects estimates in R&D Expenditures
(1) (2) (3) (4) (5) (6) (7) Output 0.289* 0.242* −0.200* −0.455* −0.625* −0.257 −0.208
(0.024) (0.030) (0.081) (0.187) (0.194) (0.206) (0.203 Output × Patent Effectiveness 1.211* 1.034* 0.184 0.262 (0.200) (0.199) (0.235) (0.238) Patent Effectiveness −6.050* −4.520* 1.173 0.552 (1.783) (1.780) (2.020) (2.046) Output × Obsolescence 0.218* 0.152* 0.117 0.100 (0.059) (0.060) (0.060) (0.060) Obsolescence −0.625 −0.121 −0.073 −0.078 (0.527) (0.533) (0.535) (0.537) Output × External Financing 0.165* 0.176* (0.030) (0.032)
External Financing −1.234* −1.341*
(0.276) (0.294) Cash Flow 0.010 0.009 0.016 0.016 0.022 0.026 0.025
(0.017) (0.017) (0.017) (0.017) (0.017) (0.017) (0.017) Cash Flow lag 0.107* 0.098* 0.105* 0.105* 0.111* 0.115* 0.119* (0.017) (0.016) (0.016) (0.016) (0.016) (0.016) (0.017) Total Assets 0.286* 0.289* 0.283* 0.286* 0.279* 0.274* 0.273* (0.019) (0.020) (0.019) (0.020) (0.019) (0.019) (0.019) Total Assets lag 0.143* 0.148* 0.143* 0.145* 0.141* 0.138* 0.135* (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) Total Liabilities 0.087* 0.070* 0.074* 0.070* 0.074* 0.077* 0.081* (0.019) (0.019) (0.019) (0.019) (0.019) (0.019) (0.019) Total Liabilities lag −0.123* −0.128* −0.125* −0.128* −0.125* −0.122* −0.119* (0.014) (0.014) (0.014) (0.014) (0.014) (0.014) (0.014) LT Debt −0.038* −0.036* −0.036* −0.036* −0.036* −0.035* −0.036* (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) LT Debt lag 0.001 0.001 0.001 0.001 0.001 0.002 0.002 (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) ST Debt −0.002 −0.003 −0.002 −0.002 −0.002 −0.002 −0.002 (0.006) (0.006) (0.005) (0.006) (0.005) (0.005) (0.005) ST Debt lag 0.002 0.002 0.002 0.002 0.002 0.002 0.002 (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) Capital Stock 0.094* 0.096* 0.101* 0.101* 0.105* 0.108* 0.107* (0.017) (0.016) (0.016) (0.016) (0.016) (0.016) (0.016) Capital Stock lag −0.008 −0.009 −0.006 −0.006 −0.003 −0.002 −0.001
(0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) C −3.656* −2.917* −1.270 −0.575 −1.082 −3.049 −3.640* (0.258) (0.317) (0.754) (1.640) (1.721) (1.825) (1.797) Year fixed effects No Yes Yes Yes Yes Yes No N (observations) 78,500 78,500 78,500 78,500 78,500 78,500 78,500 R2 0.24 0.24 0.29 0.26 0.30 0.32 0.32
6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS Table 6.3: Fixed Effects estimates for years 1991 till 1997
(1) (2) (3) (4) (5) Output −0.642* −1.193* −0.973* 0.308 −0.124
(0.187) (0.432) (0.434) (0.464) (0.440) Output × Patent Effectiveness 2.967* 2.690* 1.514* 0.748
(0.511) (0.584) (0.620) (0.584) Patent Effectiveness
Output × Obsolescence 0.502* 0.136 −0.198 −0.001 (0.135) (0.155) (0.158) (0.146) Obsolescence
Output × External Financing 0.663* 0.610* (0.088) (0.085) External Financing
Cash Flow −0.127* −0.127* −0.126* −0.116* −0.122* (0.026) (0.026) (0.026) (0.026) (0.026) Cash Flow lag −0.032 −0.033 −0.031 −0.019 −0.024
(0.020) (0.020) (0.020) (0.020) (0.020) Total Assets 0.292* 0.291* 0.292* 0.288* 0.290* (0.035) (0.035) (0.035) (0.035) (0.035) Total Assets lag 0.137* 0.140* 0.136* 0.127* 0.124* (0.022) (0.022) (0.022) (0.022) (0.022) Total Liabilities 0.111* 0.112* 0.111* 0.114* 0.112* (0.033) (0.033) (0.033) (0.033) (0.033) Total Liabilities lag −0.116* −0.119* −0.117* −0.105* −0.104* (0.020) (0.020) (0.020) (0.020) (0.020) LT Debt −0.043* −0.043* −0.043* −0.043* −0.043* (0.007) (0.007) (0.007) (0.007) (0.007) LT Debt lag 0.004 0.004 0.004 0.004 0.004 (0.007) (0.007) (0.007) (0.007) (0.007) ST Debt −0.016* −0.017* −0.016* −0.017* −0.016 (0.007) (0.007) (0.007) (0.007) (0.007) ST Debt lag −0.014* −0.014* −0.014* −0.015* −0.015* (0.006) (0.006) (0.006) (0.006) (0.006) Capital Stock 0.093* 0.094* 0.093* 0.094* 0.093* (0.032) (0.033) (0.032) (0.032) (0.032) Capital Stock lag −0.017 −0.017 −0.016 −0.018 −0.019
(0.022) (0.022) (0.022) (0.022) (0.022) Year fixed effects Yes Yes Yes Yes No N (observations) 22,090 22,090 22,090 22,090 22,090
R2 0.11 0.16 0.12 0.06 0.07
6.2. RANDOM EFFECTS ESTIMATOR CHAPTER 6. RESULTS
Table 6.4: Fixed Effects estimates including number of employees and Sales
(1) (2) (3) (4) (5) (6) (7) Output 0.269* 0.213* −0.335* −0.093 −0.039 0.260 0.300
(0.028) (0.035) (0.099) (0.242) (0.244) (0.248) (0.238) Output × Patent Effectiveness 1.483* 1.625* 0.865* 0.968* (0.251) (0.276) (0.339) (0.336) Output × Obsolescence 0.096 −0.109 −0.136 −0.154
(0.074) (0.083) (0.081) (0.077) Output × External Financing 0.179* 0.185* (0.043) (0.042) Employees 0.355* 0.370* 0.369* 0.371* 0.368* 0.370* 0.358* (0.053) (0.053) (0.053) (0.053) (0.053) (0.053) (0.053) Sales 0.041* 0.046* 0.043* 0.046* 0.043* 0.043* 0.038* (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015)
Year fixed effects No Yes Yes Yes Yes Yes No N (observations) 78,500 78,500 78,500 78,500 78,500 78,500 78,500 R2 0.25 0.26 0.23 0.27 0.20 0.17 0.17
Chapter 7
Industry groups
In this section we investigate whether procyclicality differs across major industry groups. The data used in this research exist solely of firms from the manufacturing industry. However the manufacturing industry can be further divided into a wide range of smaller industries. Each firm in the data set is appointed a 4 digit Standard Industrial Classi-fication (SIC) code. These codes are a classiClassi-fication system established by the United States and are used to identify the primary business of a firm. The 4 digit SIC codes can be categorized into progressively broader industry classifications by removing their last digit (Wikipedia, 2016).
For instance SIC code 2843 (surface active agents) belongs to industry group 284 (Soap, detergents, and cleaning preparations) which belongs to major group ’Chemicals and Allied products’ which is part of the division Manufacturing (SIC codes 2000-3999).
Table 7.1: Major Groups
SIC major group N
20 Food and Kindred Products 5,702 28 Chemicals and Allied products. 13,885 35 Industrial and Commercial Machinery and Computer Equipment. 14,040 36 Electronic and other Electrical Equipment and Components. 15,714 38 Measuring, Analyzing, and Controlling instruments. 12,807 In Table 7.1 the major industry groups are listed at which we will take a closer look. These groups have the most observations of the 20 major groups in our sample. Combined they represent over half of our estimation sample.
7.1. DUMMY ESTIMATION CHAPTER 7. INDUSTRY GROUPS
7.1
Dummy estimation
If we assume that only the variables Output × P atef f and Output × Obs differ be-tween industries, then one way to investigate if the interactions of patent effectiveness and obsolescence differences across industry groups is by interacting them with industry dummies. For each group a dummy is interacted with the variables P atEf f × Ouput and Obsolescence × Output. Basically the following terms are added to Equation 3.2:
β20Xit× P atEf fi× Ind20+ ... + β38Xit× P atEf fi× Ind38
+θ20Xit× Obsi× Ind20+ ... + θ38Xit× Obsi× Ind38
This equation is then estimated using fixed effects. The results of this estimation can be seen in Table 7.2.
There are 3 equations estimated. Column one only includes interaction terms with patent effectiveness (Pateff), the second column presents the estimation results when adding only industry interactions with obsolescence (Obs) and in the third column the full model is estimated with both sets of interactions. Table 7.2 does not show the estimates for the control variables but they were included in the estimation. It can be seen that coefficients do seem to differ across industry groups. To see if this is actually true we conduct an F-test for joint significance of the interaction slopes for each of the three estimations.
(1) H0 : β20= ... = β38 = 0
(2) H0 : θ20 = ... = θ38= 0
(3) H0 : β20= ... = β38 = θ20 = ... = θ38= 0
we find that respectively F (5, 8212) = 10.23, F (5, 8212) = 16.28, F (5, 8212) = 6.75. In each case the null is rejected. From the results in column 3 it can for instance be seen that in industry group 20 (food and kindred products) and 38 (measuring instruments) patent effectiveness has a negative effect on the procyclicality of R&D investment. Which is opposite of what is found grouping all industries together. Furthermore looking at obsolescence it can be seen that in some industry groups it has a positive effect (groups 20, 36, and 38) while in other groups it has a negative effect (groups 28 and 36).
7.2
Estimation per group
Another way to investigate whether differences in cyclicality exists between different industry groups is by simply estimation Equation 3.2 for each group separately. In this
7.2. ESTIMATION PER GROUP CHAPTER 7. INDUSTRY GROUPS Table 7.2: Fixed Effects estimates with dummy interactions
(1) (2) (3) Output −0.566* −0.857 −2.462*
(0.282) (0.447) (0.711) Output × PatEff 1.444* 2.477* (0.638) (0.819) Output × PatEff × Ind20 −7.304* −2.672
(1.709) (8.549) Output × PatEff × Ind28 0.889* 1.269
(0.405) (2.931) Output × PatEff × Ind35 0.340 2.008
(0.301) (1.161) Output × PatEff × Ind36 1.118* 1.089
(0.478) (2.602) Output × PatEff × Ind38 0.828* −2.542*
(0.359) (1.246) Output × Obs 0.313 0.520* (0.166) (0.225) Output × Obs × Ind20 −0.862* −0.496
(0.183) (0.916) Output × Obs × Ind28 0.275* −0.179
(0.0724) (0.639) Output × Obs × Ind35 −0.004 −0.326
(0.058) (0.188) Output × Obs × Ind36 0.051 0.028
(0.050) (0.281) Output × Obs × Ind38 0.262* 0.620*
(0.060) (0.230)
Year fixed effects Yes Yes Yes N (observations) 78,500 78,500 78,500
R2 0.13 0.12 0.13
∗p < 0.05
case we assume that all variables used in our analysis can possibly vary per industry group.
We estimated the full model of Equation 3.2 for each industry group mentioned in Table 7.1. However with the exception of group 38 no significant coefficients for Output × P atef f or Output × Obs were found. Therefore only the results for group 38 are reported in Table 7.3. There it can be seen that for industry group 38 there is a significant positive effect of obsolescence on the cyclicality of R&D. Something we did not find in Section 6.1. To investigate whether this coefficient truly differs from the full population we implement
7.2. ESTIMATION PER GROUP CHAPTER 7. INDUSTRY GROUPS Table 7.3: Fixed Effects estimates group 38
Output −4.234* (0.1.734) Output × PatEff 5.336 (0.3.919) Output × Obs 0.902* (0.348) N (observations) 12,410 R2 0.21 ∗p < 0.05
a structural break or Chow test (Cameron and Trivedi, 2005). The statistic is: (Sc− (S1+ S2))/k
(S1+ S2)/(N1+ N2− 2k)
with Sc is the sum of squared errors of the unrestricted model. S1 and S2 are the
sum of squared errors for respectively the model estimated only using observations in industry group 38, and the model estimated without observations from group 38. k are the number of variables used in the model. The test statistic has a F distribution with k and (N1+ N2− 2k) degrees of freedom. For the R&D model we get F (43, 78414) = 18.435
and we reject the null that we can pool the data. However the Chow test requires that the variances of the errors are equal across samples. In our case this is not true and how this exactly affects the Chow test is unclear. A study done by Toyoda (1974) states that ”First, the Chow test is well behaved even under heteroscedasticity as long as at least one of two sample sizes is very large.”
Concluding, we see that how patent protection levels and obsolescence influences the procyclicality of R&D differs across industry groups. Especially using dummies we can see a lot of variation. It could therefore be that the consistency of parameters across groups that fixed effects requires might not be realistic.
Chapter 8
Conclusion
This research expands upon the results found in the paper An Emperical Examination of the Procyclicality of R&D Investment and Innovation written by Fabrizio and Tsolmon, by using different estimation techniques. Namely a fixed effects and random effects model instead of a first difference model. The results were compared to the results of Fabrizio and Tsolmon (2014). Furthermore we take a closer look at differences between groups of industries and how cyclicality differs.
Contrary to the research done by Fabrizio and Tsolmon (2014), using fixed effects estimation, we found no evidence that firms in industries with a faster rate obsoles-cence behave more procyclical. Secondly using fixed effects this research finds that firms in industries with a higher level of patent effectiveness are more procyclical instead of less. Using random effects estimation we found that faster obsolescence does increase the procyclicality of R&D investment, just as in Fabrizio and Tsolmon (2014). For patent effectiveness it found the same results as the fixed effects model, firms operating in in-dustries with higher patent protection levels are less procyclical. However, a Hausman test indicated that fixed effects are present and therefore our random effects estimates are probably biased.
Additionally, looking at our data’s largest industry groups, we find evidence that the influence of patent effectiveness and obsolescence on the procyclicality of R&D does differ across industry groups. Indicating that the data might not be poolable.
Bibliography
Aghion, P., Askenazy, P., Berman, N., Cette, G., and Eymard, L. (2012). Credit con-straints and the cyclicality of r&d investment: Evidence from france. Journal of the European Economic Association, 10(5):1001–1024.
Barlevy, G. (2007). On the cyclicality of research and development. The American Economic Review, 97(4):1131–1164.
Caballero, R. J. (2010). Creative Destruction. Palgrave Macmillan UK.
Cameron, A. C. and Trivedi, P. K. (2005). Microeconometrics: methods and applications. Cambridge university press.
Fabrizio and Tsolmon (2014). An emperical examination of the procyclicality of r&d investemnt and innovation. The Review of Economics and Statistics, 96(4):662–675. Francois, P. and Lloyd-Ellis, H. (2003). Animal spirits through creative destruction. The
American Economic Review, 93(3):530–550.
Hall, H. B., Griliches, Z., and Hausman, A. J. (1986). Patents and r&d: Is there a lag? International Economic Review, pages 265–283.
Hud, M. and Rammer, C. (2015). Innovation budgeting over the business cycle and innovation performance. ZEW-Centre for European Economic Research Discussion Paper, (15-030).
Schumpeter, J. (1942). Capitalism, Socialism, and Democrazy. Harper.
Shleifer, A. (1986). Implementation cycles. The Journal of Political Economy, 94(6):1163– 1190.
Toyoda, T. (1974). Use of the chow test under heteroscedasticity. Econometrica, 42(3):601–608.
Wikipedia (2016). Standard industrial classification — wikipedia, the free encyclopedia. [Online; accessed 30-November-2016].
