Using the propensity score matching method to analyze the effect of R&D subsidies

(1)

Universiteit van Amsterdam 17-03-2014 Joep Hoefsloot Matriculation number: 6168752 Thesis supervisor: Maximilian O. Hoyer

Using the Propensity Score Matching

method to analyze the effect of R&D

subsidies

1. Introduction

In the last decades, technology has improved quicker than ever. Together with this rise in technology and its importance, technological competition between firms and countries got fiercer. Countries leading the technological innovation like the U.S., Germany and China are doing everything they can to maintain their competitive advantage. Most countries have been increasing their investments into technological innovation to keep up with this race to expand their economy and boost its growth. A variety of instruments are present for governments to boost technological change. Examples are tax cuts 1, subsidies, national R&D entities, patents, forming R&D consortia between universities and settled companies, and promoting technical studies. Subsidies however, are the most questionable. A lot of studies have been conducted on the effectiveness of R&D subsidies. However, how effective R&D subsidies really are is still unclear. That is why I will try to cover all important aspects of this

(2)

subject, look at the post crisis situation, bring new empirical results and suggest some solutions to understand the problems concerning R&D subsidies a bit better.

The standard argument for R&D subsidies is that there is some kind of market failure which leads to underinvestment in R&D in the private sector (Arrow (1962); Nelson (1959)). Most of the underinvestment in private R&D stems from the fact that the profits from R&D are often hard to appropriate for firms. Profit here means more than just the profits earned by the firm that invests in R&D, but also the benefit other people get from the innovation. Thus, the social benefits from R&D projects are higher than the private benefits. Because firms can only appropriate private benefits, costs of an R&D project can be small compared to the social benefits but might still be too big for the private benefits meaning the project will not be carried out. The result is that underinvestment occurs and the socially optimal level is not reached.

How will a subsidy stimulate current and future R&D investments? Following Lach (2002), R&D subsidies lower the private cost of R&D projects which could result in non-profitable projects being turned into profitable projects (thus more likely to be undertaken). R&D subsidies can also account for building- and upgrading research facilities which lower the fixed costs of other and/or future projects, increasing the probability of those projects being undertaken. Lastly, the knowledge gathered during a R&D project may also spill-over to other current and future R&D projects and so enlarge the probability for those projects being undertaken.

In general, there are three main questions to be asked about R&D subsidies to determine its effectiveness:

1. Is the subsidy selection process efficient?

2. How does a subsidy influence the decisions made by firms on private R&D investments? 3. What is the final innovation effect of R&D subsidies?

For the first question we need specific data on the selection process of subsidies. For the third question we need additional data on for example the number of new products released on the market per firm. Because I do not have data available to answer question 1 and 3, I will focus my empirical study solely on question 2. A specific explanation about the dataset will be done later.

The paper will continue as follows: first, in the Literature Review(2), I will explain the theory behind R&D subsidies and the results of other studies over the last decades. Then I will discuss the Data(3) available to me. Following up I will show the Method(4) used to get the Results(5). And lastly I will discuss the results and some important facts involved in this study combined with my conclusions in the Conclusion and Discussion(6).

(3)

2. Literature review

As explained in the introduction, subsidies boost innovation spending because they are ought to be seen as a complement. But is this really the case? Not necessarily, the phenomenon “Crowding-out” is the negative effect which can be the result of a firms decision on private R&D investments after the firm has been granted a subsidy. So to answer the questions 1 and 2 from part 1, Introduction, we need to investigate the reasons for- and facts about “Crowding-out”.

Crowding-out is a negative substitution effect which occurs when a firm receives a subsidy and is not using it as a complement for more private R&D investments, but (partly) as a substitute for its own private R&D investments. There are several reasons why firms are inclined to use the subsidy as substitute instead of a complement. The most obvious reason is that a subsidy is a cheaper, almost free (except for the administration) way of getting money compared to raising money in the capital market over which you pay interest. So instead of planning to do more private investments in R&D with the received subsidy, the firm can choose to keep the R&D investments equally high as before the subsidy, and use the subsidy as a substitute for their own R&D investments. What might help against this crowding-out motive is including a clause with each subsidy to make sure a firm does not change its R&D investment portfolio after getting a R&D subsidy.

Another way through which crowding-out can occur is through the effect on the price of in-elastically supplied R&D inputs (David et al. (2000). Suppose a given subsidy turns an unprofitable project into a profitable project. Then if some inputs, for example workers, are very expensive, the firm may not be able to continue all ongoing R&D projects, plus this new profitable project. This means that the firm will have to shut down one of its proceeding projects to afford the new project. This also is crowding-out of private R&D investments (more specifically: crowding-crowding-out of R&D projects). On the other hand, the giving party, for example the government who hands out the subsidies, can also be a reason for crowding-out. Bureaucrats may be under strong pressure to avoid the appearance of “wasting” public funds and therefore, may tend to fund projects with higher success probabilities and with clearly identifiable results, i.e. projects that are likely to have high private rates of returns (Lach (2002)). These projects with a high success probability will be able to be financed anyways, by private capital or the capital market, which means the subsidy might be superfluous and crowd-out private R&D investments. This would of course not be the case if the superfluous capital for a certain project would be invested in other projects.

Knowing that crowding-out is in theory likely to occur, we have to test this effect empirically. Wallsten (2000) uses simultaneous equation systems and instrumental variables to control for endogeneity to find out that subsidies crowd out private R&D expenditures dollar for dollar after firms got the award of a Small Business Innovation Research grant in the U.S.. Busom (1999) finds in her

(4)

firms. On the other hand, Klette and Moen (1997), Lach (2002) and Hussinger (2003) find much less disapproving results. Klette and Moen (1997) show that there is not much evidence for crowding-out in their sample of high-technology Norwegian firms. They also conclude that the R&D subsidies were successfully targeted at firms that have significantly expanded their R&D expenditures. Lach (2002) finds that subsidies have a positive total effect on small firms after estimating the relative increase in R&D expenditures of subsidized firms versus non-subsidized firms using panel data from a sample of Israeli firms. The article states also that positive total effect diminishes when looking at larger firms. Hussinger (2003) shows a positive effect of subsidies on the R&D spending of firms using a semi-parametric model of sample selection on a sample of German manufacturing firms. González et al. (2005) estimate that for Spanish firms, subsidies increase total R&D expenditures by slightly more than their amount.

Now we know that there is a lot more research the to be done about the ‘Crowding-out’ effect, we can look at more specific questions on this effect. From the studies just mentioned, Lach (2002) and González et al. (2005) both point out that most subsidies are going to projects that would have been undertaken also without the R&D subsidy, which supports the earlier argument that bureaucrats may be under pressure avoiding spilling public funds. This indicates that there should be done more research on which projects to subsidize and the subsidizing parties should be more careful deciding whom to grant R&D subsidies.

Until Busom (1999) stated some major arguments against simple regression models, these models were the main method used by researchers to estimate the effect of subsidies on private R&D investments. First, she explains the problem of endogeneity of public funding. “In order for a firm to receive public funding it must apply for funding, and the public agency may or may not award it, given firm and project characteristics. This makes public funding an endogenous variable, and its inclusion in a linear regression will cause inconsistent estimates if it happens to be correlated with the error term” Busom, (1999). Since it is likely that unobservable factors determine both public funding and private R&D investment decisions, inconsistent estimates are likely to occur. Secondly, Busom states that “the conclusion that should be drawn from finding a negative relationship between public and private R&D expenditure in former studies is not clear, as the public agency may choose precisely to finance more heavily R&D projects with higher spillover potential (for instance, those involving basic research or higher appropriability difficulties), where incentives for private funding may be small. In that case, there is no reason to expect positive or higher private spending by the recipient firm, unless the project triggers additional applied and development research generating private returns”. This argument tells us that if public funding is involved with little private funding (resulting in a non-significant or negative coefficient in the regression), this does not mean that crowding-out out effects are necessarily present.

(5)

To avoid the problem of endogeneity brought up by Busom, studies after the year 1999 started to use different estimation methods. Most articles from then on used some kind of matching method between a treatment group and a control group (e.g. Almus and Czarnitzki (2003); Hussinger (2003); González et al. (2005); González and Pazó (2008); Busom (1999)). Along with the new estimation methods, most importantly matching methods, other datasets are being used. These datasets are easy to use for matching treatment groups with control groups and have a lot of observations and variables which leads to better estimations. They also give the opportunity to match. These kind of datasets also have their drawbacks, which will be discussed in section 6, Conclusion and Discussion.

In my empirical research the focus will be on the overall effect of subsidies on the private R&D investments using Propensity Score Matching. This is a straight forward matching method that is easily applied with the STATA software (version 12). Subsidized firms will be matched with non-subsidized firms and I will then come up with an average difference in R&D intensity (R&D investments/ total turnover) to see if the R&D subsidies have a positive or negative overall effect on private R&D investments. R&D intensity is the most convenient available variable that measures the amount of private R&D investments and is commonly used in this field of research (see: Busom (1999); Czarnitzki et al. (2007); Lööf and Heshmati (2005); Almus and Czarnitzki (2003)). This method will be applied onto three years: 2001, 2003 and 2009. The first two years are recent years with the most data available.

For the years 2001 and 2003, my focus will be on verifying whether R&D subsidies have a positive overall effect on private R&D investments. This is an essential question for governments to decide whether they should grant R&D subsidies or not. As stated in this section already, there is a lot of counterfactual data on this. Many studies show (partly) crowding-out, whereas others do not. In case of the main recent articles discussing the crowding-out in Germany (Almus and Czarnitzki (2003); Hussinger (2003), significant positive effects for R&D subsidies were found. Therefore my first hypothesis is as follows: R&D subsidies will have a significant positive overall effect on private R&D investments.

On top of that, I have added the year 2009. This year might provide us with some insights in how the financial crisis influences the private RD investment decisions of innovating firms. I expect investors to be less risk-taking after the financial crisis. It will be harder for innovating firms to attract money from the capital market. Together with the fact that the financial crisis has most likely lowered the annual turnover, firms are expected to be more financial constrained. Therefore firms are likely to loosen up their financial constraints by means of using the R&D subsidy as a substitute. From this I get my second hypothesis: the effect of the R&D subsidies on the private R&D investments the years closely after the financial crisis will be smaller compared to the years before the financial crisis.

(6)

Lastly, I will examine if the size of a firm is correlated with the effect of R&D subsidies on the private R&D investments and verify if the positive effects of R&D subsidies really diminish with the increase of firm size as Lach (2002) states.

3. Data

The data used for this research comes from an anonymized survey from the Mannheim Innovation Panel (MIP) that has been conducted by the Centre of European Research (ZEW2) on behalf of the German Federal Ministry for Education and Research. The MIP survey focuses on innovation activities in the business sector and is solely used for scientific purposes.

The survey has data available from 1993 to 2010 for the Manufacturing and Mining sector. I will use the data from the manufacturing and mining industry in the years 2001, 2005 and 2009 as explained in the Literature Review. Since the 3 separate years only partly consist of the same firms (i.e. if I would use 3 consecutive years, less than 10 % of the firms would take part in all of these 3 years), I will have to use the data as separate cross-sections.

The survey included the variable “Filter” to filter non-innovating firms out of the dataset. The definition if a firm is innovating or not is based on the Oslo-Manual guidelines (see Eurostat and OECD (1997)) which says that “innovators are defined as firms which have introduced at least one product or process innovation in the 3 most recent years”. After filtering the data for non-innovating firms, the datasets consist of approximately 3000 observations each on innovating firms in Germany from which approximately 90% received R&D subsidies.

Most of the former research has been done on particular R&D schemes which cannot control for possible other R&D subsidy programs the firms might get and could lead to biased results. For example, if a firm gets another subsidy next to the subsidy which is studied, the data might show that the private R&D investments went up significantly, yet not necessarily because of the studied subsidy. This might lead to biased results. A great advantage of this dataset is that it covers all R&D subsidy schemes so I can look at the collective average effect of all R&D subsidy programs. Specific effects of different types of subsidies are left for other researches.

(7)

4. Method

Model: Propensity Score Matching

The easiest way next to the Ordinary Least Squares estimation (OLS) to find the effect of a subsidy would be to look at the Average Treatment Effect (ATE):

ATE = E(Δ) = E(y1|x, D = 1) - E(y0|x, D = 0)

Where D = 1 for treated observations and D = 0 for control observations. y1 Stands for outcome being treated and y0 for outcome if not being treated. x being the firm its characteristics (e.g. independent variables affecting the likelihood to get a subsidy).

But since this is not a random experiment but an observational study, the ATE may be biased if the control observations and the treated observations are not similar. This is why I use the Average Treatment Effect on the Treated (ATET).

The ATET measure is the difference between the outcomes of the treated and the outcomes of the treated observations had they not been treated.

ATET = E(Δ|D = 1) = E(y1|x, D = 1) - E(y0|x, D = 1)

As you can see the last term is counterfactual. We cannot know the outcomes of the treated observations had they not been treated. We can however use the Propensity Score Matching (PSM) to estimate this last term.

The propensity score (p(x)) is the probability a firm gets the treatment (subsidy). The propensity score is estimated by a probit/ logit model with pre-treatment characteristics x that may affect the likelihood of getting the treatment.

P(x) = prob(D = 1|x) = E(D|x)

After estimating p(x) for both the treated and the control group we can match them on several ways. Nearest Neighbor Method is the most commonly used but since I don’t have a lot of control observations, the Kernel Matching Method seems to be the better choice.

 Nearest Neighbor Method: match the treated observations with the control observations with the nearest p(x).

 Kernel Method: Take all control observations into account and weigh them according to how close their p(x) is to the treated observation.

(8)

After matching on propensity scores, I can compare the outcomes of the treated and control observations the following way:

ATET = E(Δ|p(x), D = 1) = E(y1|p(x), D = 1) - E(y0|p(x), D = 0) There are two ways to improve the performance of PSM:

 Common Support: With common support the matching is restricted to the common range of the propensity scores of the treated group and the control group.

 Bootstrapping: “Each bootstrap draw includes the re-estimation of the results, including the first steps of the estimation (propensity score, common support, etc.). Repeating the bootstrapping N times leads to N estimated average treatment on the treated effects. The distribution of these means approximate the sampling distribution (and thus the standard error) of the population mean. ” (Caliendo and Kopeinig (2005). Since we take the average of all the standard errors of the N replications, we will end up with a smaller standard error.

As noted in the literature review, PSM solves the selection bias, but has some strong assumptions. I will now look at the assumptions concerning PSM brought up in Caliendo and Kopeinig (2005) and Bryson et al. (2002).

1. Conditional independence assumption: if one can control for observable differences in characteristics between the treated and the non-treated group, the outcome that would result in the absence of the treatment is the same in both cases.

2. Explanatory assumption: Only variables that influence simultaneously the participation decision and the outcome variable should be used. (the more the variable determines both decision and outcome by the same extent, the more precise the estimates will be).

3. Exogeneity assumption: Only variables that are unaffected by participation should be included (to ensure this, these variables should be fixed or measured before participation). Two more notes on choosing the right PSM variables: important for choosing the right variable is that omitting important variables might seriously increase the bias in resulting estimates according to Heckman et al. (1997). Secondly, Heckman, LaLonde and Smith (1999) note that data for the two groups should come from the same source.

Assumption (1) implies that selection is solely based on observable characteristics and that all variables that influence treatment assignment and potential outcomes simultaneously are observed by the researcher. Therefore, any differences between the treated and the non-treated can be attributed to the effect of the program. This is of course a too strong assumption to be true in most of the studies. If assumption (1) is met, the problem of ‘too good’ data could arise. If p(x)=0 or p(x)=1 for the same values of x, we could not use PSM since matching would not be possible. Assumption (2) tells us that

(9)

the more the variable explains, the better. Assumption (3) keeps endogeneity between the treatment and the variables out. In this study I will assume assumption 1 is met.

Variables:

Outcome variable: R&D intensity (variable name: rdint). Total R&D expenditure as a fraction of turnover (if larger than .15, it is truncated to .153).

Treatment variable (variable name: treat): Public subsidies of the federal states, from the federal government (ministries)/ EU. This variable only explains whether a firm got a subsidy in a certain year, not what amount the firm has gotten.

Next I have to look whether the variables I chose (to determine the probability of being in the treated group) for my research (partly) fulfill assumptions (2) and/ or (3). This will be done according to economic theory. In part 5, Results, I will only use the variables which fulfill the assumptions. If those chosen variables are significant will also be seen in the Results.

Yes means the assumption is met, No means the assumption is not met.

Note: for assumption (2), it is not entirely possible to determine if the variable influences the participation decision since we do not know what kind of firms apply for a subsidy/ what kind of firms are picked out for the subsidy programs. Though, we can argue those facts. I will give an example of the possible line of thought with the first variable in table 1, startup. From then on, only explanations show which are non-trivial.

(10)

Tabel 1: Variables and their assumption fulfillments

Variable Description Assumption 2 Assumption 3

startup founded within the last 3 years or not

Yes (for example, subsidy granting parties might tend to grant R&D subsidies mostly to well established firms and new firms are likely to have higher R&D intensity ratios than older firms)

Yes (the fact that a firm is founded in the last 3 years or not cannot be affected by subsidies after the founding date)

merger merger or not Yes Yes

smarket main selling market: international/ national-local/ nationwide

Yes Yes

turno turnover No No

ftempl full time employees Yes Yes,

the influence of a possible rise in R&D employees won’t be significant, Almus and Czarnitzki (2003) inno who develops product

innovations? Firm/ firm with other firm/ mostly other firm

Yes Yes

newp market novelties or not Yes No

rdactiv continuous R&D activities/ no R&D activities/ occasional R&D activities

Yes No

rdcoop innovation related cooperation with other firms or public research establishments or not

(11)

empl # employees: 0-50 -250->250

Yes Yes, see ftempl for

explanation

lp labor productivity:

turnover/ # employees

Yes Yes, same logic as seen

with ftempl exs export rate: turnover

from abroad/turnover. (if larger than .85, truncated to .85)

Yes Yes

inshare total innovation

expenditure/turnover. (if larger than .355, truncated to .35) Yes No, strong endogeneity/correlation rdintp R&D employees/employees. (if larger than .15, truncated to .15)

Yes No,

since R&D employees is part of the equation, the reasoning as witch ftempl cannot be applied. patent patents in the last 3 years

or not

Yes Yes

5. Results

First I will look at the overall results of R&D subsidies on the R&D intensity for the year 2001 and 2003 and show the method used in detail. After this I will do the same for the year 2009, with less details. Then I will look for any differences in overall effects of R&D subsidies on the R&D intensity between small and large firms.

Before using the datasets I filter on whether firms are innovative or not by filtering on the variable ‘Filter’. This is because non-innovators, which have an R&D intensity of 0 do not qualify for subsidies and thus should have a propensity score of 0. However, because they might have similar x variables as firms that do innovate the propensity score may be different from 0 when computing it, this will bias our results. After filtering, 1000 to 4000 observations remain depending on the year. Only 300 to 600 will be used (depending on the year) since only those observations have values for the variables we based the probit model on.

(12)

I first use OLS to test the significance of the treatment variable. Then I will test the significance of the variables which are used for the probit model to calculate p(x).

To check whether treatment is a significant explanatory variable to the R&D intensity, I regress the R&D intensity on treatment (same as t-test) and I got a P-value of 0.001, which means the treatment variable is significant.

When we control for the variables startup merger smarket inno empl lp exs patent, it gives the following results:

Tabel 2: OLS on rdint (R&D Intensity), 2001

Variable Coef Std. Err. T P>|t| [95% Conf. Interval]

treat 0.0189217 .0162382 1.17 0.245 -.0130955 .0509389 startup -0.0055586 .0215378 -0.26 0.797 -.0480251 .0369080 merger -0.0117763 .0110942 -1.06 0.290 -.0336509 .0100983 smarket 0.0069684 .0047248 1.47 0.142 -.0023476 .0162844 inno -.0075557 .0063322 -1.19 0.234 -.0200411 .0049297 empl -.0278707 .0048214 -5.78 0.000 -.0373772 -.0183642 lp -.0338750 .0245068 -1.38 0.168 -.0821954 .0144455 exs .0118457 .0145593 0.81 0.417 -.0168611 .0405525 patent .0400195 .0070104 5.71 0.000 .0261970 .0538420 _cons .0537839 .0245191 2.19 0.029 .0054391 .1021287

As you can see the variables have very high P-values. To get lower P-values I eliminated one variable, smarket, to get rid of plausible Multicollinearity (Multicollinearity biases the results and occurs when if two or more independent variables in a regression are highly correlated with each other). Smarket is likely to be highly correlated with exs because both variables tell us something about how much firms trade (inter)nationally. When doing the regression again you can that the variables have much lower p-values. Secondly, I also deleted the variable Merger because Bryson et al. (2002) state two reasons why over-parameterized models should be avoided. First, it may be the case that including unrelated variables in the participation model worsen the (common) support problem. Secondly, although the inclusion of non-significant variables will not bias the estimates or make them inconsistent, it can increase their variance.

(13)

Tabel 3: OLS on rdint (R&D Intensity) without Merger and Smarket, 2001

Variable Coef. Std. Err. T P>|t| [95% Conf. Interval]

treat .0232099 .0141442 1.64 0.102 -.0046424 .0510622 startup -.0103421 .0138995 -0.74 0.458 -.0377125 .0170282 inno -.0081599 .005636 -1.45 0.149 -.0192581 .0029384 empl -.0274393 .0041514 -6.61 0.000 -.0356141 -.0192646 lp -.0402231 .0215228 -1.87 0.063 -.0826051 .0021588 exs .0301023 .0118029 2.55 0.011 .0068603 .0533442 patent .0357068 .0058808 6.07 0.000 .0241266 .047287 _cons .0684683 .0187328 3.65 0.000 .0315802 .1053564

When checking the significance of the variables on the treatment (see table 4), we see that the variables which explained the outcome variable the best compared to the other variables, explain the treatment variable relatively the worst, and the other way around. This shows that assumption (2) of PSM is hard to meet, but this will not worsen the performance of PSM as we will see later on.

Tabel 4: OLS on treat (Subsidy), 2001

Variable Coef. Std. Err. t P>|t| [95% Conf. Interval]

startup -.1388775 .0550199 -2.52 0.012 -.2471681 -.030587 inno -.0712858 .0221535 -3.22 0.001 -.1148886 -.0276831 empl .0017321 .0162491 0.11 0.915 -.0302495 .0337136 lp -.061113 .0869423 -0.70 0.483 -.2322333 .1100074 exs .0734067 .0468373 1.57 0.118 -.0187787 .1655921 patent -.0083966 .0233621 -0.36 0.720 -.054378 .0375847 _cons 1.052862 .0445501 23.63 0.000 .9651787 1.140546

After using STATA to create a probit model using the variables as stated in table 4, all propensity scores in region of common support are between 0.762 and 0.998 with a mean of 0.964 and a standard deviation of 0.039. The mean of 0.964 shows the little amount of control observations. Fortunately, the other two years have significantly more control observations. See Table 5 for specifications of the propensity scores of the year 2001.

(14)

Table 5: Propensity scores, 2001

Percentiles Prop. Score Smallest

1% .7834963 .762392 5% .8956052 .7724877 10% .9084601 .7732039 25% .9613548 .7755101 50% .9735458 Largest 75% .9874996 .9971783 90% .9938527 .9972372 95% .9953814 .9972535 99% .9968214 .9975562

Next are the results for the year 2001 applying Kernal Matching with common support and bootstrapped standard errors (50 replications).

Table 6: The Avarage Treatment Effect on the Treated (Kernal Matching), 2001

n. treat. n. contr. ATET Std. Err. T

290 10 0.028 0.004 6.276

As you can see in Table 7, the Nearest Neighbor method gives a larger standard error because we have few control observations. This is why we will use Kernal Matching in the rest of the study.

Table 7: The Avarage Treatment Effect on the Treated (Nearest Neighbor Matching), 2001

n. treat. n. contr. ATET Std. Err. t

290 10 0.028 0.009 3.042

As the tables show, the ATET is positive and highly significant at a 1% significance level with a t-score of 6.276. An ATET of 0.028 means that the R&D intensity is on average 0.028 higher for firms with R&D subsidies than for firms without R&D subsidies. This can be translated into a higher R&D intensity of 2.8 % for subsidized firms than for firms without a subsidy on average. Though, since we do not know anything about the size of the subsidies, we cannot say anything about the relative size of the ATET.

(15)

Because the OLS estimations are done on 2001, the year with the least observations (more importantly, the control observations account only for 3.4 % of the total observations), I quickly go through the OLS estimations again for the year 2003. Starting with an OLS estimation on rdint, shown in table 8. Smarket and patent are not included in the regression because those variables are not available for 2003 (and 2009). Because startup has a very high p-value and we do not want to over-parameterize, I deleted this variable from the regression, shown in table 9. Lastly, I regress the treatment variable on the independent variables, as seen in table 10. Again, the high p-values in table 10 lower the PSM performance as seen in the ATET results at the end.

Table 8: OLS on rdint (R&D Intensity), 2003

Variable Coef. Std. Err. t P> |t| [95% Conf. Interval]

treat .0409966 .0063484 6.46 0.000 .0285214 .0534719 startup .0021777 .0107055 0.20 0.839 -.0188597 .0232152 merger .0139824 .0078628 1.78 0.076 -.0014689 .0294336 inno -.0058152 .0045484 -1.28 0.202 -.0147532 .0031228 empl -.0246788 .003308 -7.46 0.000 -.0311794 -.0181781 lp -.0670697 .0175113 -3.83 0.000 -.1014811 -.0326583 exs .0271349 .0097127 2.79 0.005 .0080485 .0462213 _cons .0862013 .0103076 8.36 0.000 .0659459 .1064567

Table 9: on rdint (R&D Intensity) without startup, 2003

Variable Coef. Std. Err. t P>t [95% Conf. Interval]

treat .040974 .0063187 6.48 0.000 .0285572 .0533907 merger .01107 .0076991 1.44 0.151 -.0040592 .0261991 inno -.0055754 .0045509 -1.23 0.221 -.0145182 .0033674 empl -.0246049 .0032916 -7.48 0.000 -.031073 -.0181367 lp -.0659505 .0175085 -3.77 0.000 -.100356 -.031545 exs .0273073 .009699 2.82 0.005 .008248 .0463665 _cons .0855673 .0103059 8.30 0.000 .0653155 .1058191

(16)

Table 10: OLS on treat (Subsidy), 2001

Now I will continue with estimating the ATET for the years 2003 and 2009 using PSM. After creating a probit model using the variables as stated in table 10, all propensity scores are between 0.463/0.715 and 0.939/0.968 with a mean of 0.829/0.872 and a standard deviation of 0.075/0.051. With this estimation common support was applied. See Table 11/12 for specifications of the propensity scores. Table 11: Propensity scores, 2003 Table 12: Propensity scores, 2009

Next are the results for the year 2003/2009 applying Kernal Matching with common support and bootstrapped standard errors (50 replications).

Variable Coef. Std. Err. t P>t [95% Conf. Interval]

merger .0494362 .0553974 0.89 0.373 -.0594063 .1582786 inno -.0696081 .0327454 -2.13 0.034 -.1339448 -.0052713 empl .0387606 .0233995 1.66 0.098 -.0072138 .0847349 lp -.3816073 .1245349 -3.06 0.002 -.6262884 -.1369263 exs -.0065834 .0699154 -0.09 0.925 -.1439503 .1307835 _cons .947664 .0609345 15.55 0.000 .8279425 1.067385

Percentiles Prop. Score Smallest

1% .5652066 .4633672 5% .6799398 .4633672 10% .7315508 .481156 25% .7929798 .4842585 50% .8498734 Largest 75% .8816128 .9320878 90% .9016154 .9335482 95% .909935 .9339116 99% .9258343 .9393418

Percentiles Prop.Score Smallest

1% .7289273 .7153259 5% .7647248 .7153259 10% .8010925 .7153259 25% .8462309 .7153259 50% .875577 Largest 75% .9086827 .966597 90% .9349201 .9668441 95% .9460644 .9669965 99% .9615157 .9682932

(17)

442 87 0.043 0.005 8.496

The results of the year 2003 and 2009 show us a higher ATET (0.043 and 0.043 compared to 0.028) with relatively smaller standard errors. This means that the fact if a firm receives a subsidy on average causes an increase of 4.3% in R&D intensity. That is 1.5 % higher than in the year 2001. The smaller standard errors are likely the result of a much higher amount of observations, and more importantly, a much higher fraction of control variables compared to the year 2001 (20% and 13% compared to 3.4%).

My hypothesis about the post-financial crisis data proves to be wrong. There is no lower ATET in 2009 than in the other two years. In fact, the overall effect is higher than in 2001 and the same as in 2003. I will discuss these findings in the final section.

Now I will check whether firm size has influence on the overall effect of R&D subsidies on the private R&D investment decisions of firms. To do this I have looked at every years’ total turnover (variable: turno) its median and split the data of each year in a ‘small firm’ and ‘big firm’ section considering total turnover. I then do the same PSM as before on those two sections of each year. As seen in the tables 15 to 20, small firms have a significant higher ATET than large firms. The average rise in R&D intensity due to the fact if a firm receives a subsidy is 1.1 % higher for small firms than for big firms in the year 2001. For the years 2003 and 2009 this rise is 4.8% and 3.8% larger for small firms than for big firms. This means that the overall positive effect of R&D subsidies on the private R&D investments diminishes when firms get larger.

(18)

Table 15: Small firms ATET, 2001

151 7 0.034 0.008 4.215

Table 16: Large firms ATET, 2001

139 3 0.023 0.008 2.838

255 47 0.064 0.010 6.177

187 38 0.016 0.005 3.028

277 41 0.062 0.007 8.529

(19)

6. Conclusion and Discussion:

Using Propensity Score Matching on a German dataset I have come to the conclusion that R&D subsidies have a positive overall effect on private R&D investments. I found significant positive overall effects of subsidies on private R&D investments of 2.8%, 4.3% and 4.3% respectively for the years 2001, 2003 and 2009. This result means that when a firm receives a subsidy, on average this firm has a R&D intensity that is 2.8-4.3% higher than for firms which did not receive a subsidy. This confirms my first hypothesis that says that R&D subsidies will have a significant positive overall effect on private R&D investments and supports the findings of Almus and Czarnitzki (2003) and Hussinger (2003). With these findings it should be clear that governments should not be worried about subsidizing R&D projects. However, governments should focus more on which projects to subsidize (namely projects with high spillover effects and high social benefits instead of projects with high private benefits which might be more alluring for bureaucrats whom are under pressure of wasting public funds). Also, with the findings of my empirical research about relation between firm size and the effectiveness of R&D subsidies, it has become clear that it is more efficient to subsidize small firms. My research shows that the R&D intensity of small firms grow 1.1%, 4.8% and 3.8% more than the R&D intensity of large firms when receiving a subsidy respectively for the years 2001, 2003 and 2009. This confirms the findings of Lach (2002) that tell us that the positive effect of R&D subsidies diminishes with the increase of firm size.

Lastly, my empirical research rejects my second hypothesis that says that closely after the financial crisis the R&D subsidies will have a smaller effect on the private R&D investments than before the financial crisis. Specifically, I found that the effect of R&D subsidies in 2009 was bigger than in 2001 and equal to the effect of R&D subsidies in 2003. This result tells us that the financial crisis had no negative effect on the effectiveness of R&D subsidies on private R&D investments. This is a contra intuitive result which will need more research. An idea to improve the research on the effect of the financial crisis on the effect of R&D subsidies would be to use datasets from other countries on which the crisis might have had a more drastic impact.

González and Pazó (2008) verifies that most recent studies apply matching estimators (just like Propensity Score Matching) and points out that the samples used for those studies often only include data from firms which perform R&D and/or have no information about the amount of R&D subsidies those firms receive. The former excludes the possibility to look at an ‘inducement effect’. This effect would show us if non-innovators are likely to start making R&D expenditures when they get a R&D subsidy. The latter tells us that we do not know how much of the R&D expenditures are their own. This means that partial crowding-out cannot be assessed. Dueget (2004) emphasizes this problem by stating that the size of a subsidy is likely to have an effect on the crowding-out effect. Lastly,

(20)

variables which would allow us to consider the persistence of the innovative activity. This is called the ‘persistence effect’. So panel-data instead of cross-sectional data would be preferred to measure the ‘persistence effect’.

All drawbacks mentioned by González and Pazó (2008) apply to my study as well. But in contrast to what González and Pazó say, it might be more convenient to get an overall result on whether subsidies have an positive effect on R&D investments before questioning an ‘inducement effect’ or looking at whether full or partial crowding-out takes place. Though, the lack of measurement on the ‘persistence-effect’ with only cross-sectional data is likely to bias the outcome of the overall effect of a subsidy on private R&D investments since positive lagged effects are not observed.

Of course, more observation would give better estimations. Also with more observations it might be possible to compare the effect of subsidies on the private R&D investment decisions between different branches in the Manufacturing and Mining sector to determine which sector uses the subsidies the most efficient (i.e. where the firms private R&D investments rise the most due to R&D subsidies). As shown in section 4, Method, Propensity Score Matching has strong assumptions which are hard to meet. Since we have no data on the kind of subsidies which are granted in the survey, it is hard to find good variables for the probit model to determine the propensity a firm gets a subsidy (Propensity Scores). On the other hand, as explained in section 4, Method, ‘too good’ data could data would make it impossible to use PSM since for the same values of x, p(x)=0 and p(x)=1 would occur.

A shift in direction in this field of research might be considered since the private R&D investments/ R&D intensity as the dependent variable has been proven to have a lot of insecurities because of unsatisfying datasets combined with limits of the estimation methods. On top of that, Goolsbee (1998) notes that because of the fact that the R&D labor supply is quite inelastic, and the R&D investments are mainly just the wages of R&D employees, a significant fraction of R&D subsidies/ increased R&D spending of the firm goes directly to the wages of the scientific personnel. Together with the fact that in former studies there has only been focus on the effect of subsidies on private R&D investments, it might be more convenient to look for final effects of R&D subsidies instead, i.e. the third question from part 1, Introduction: What is the final innovation effect of R&D subsidies?

A few examples to look at the final effects of R&D subsidies more directly are: collecting data on how many patents a firm have been granted in a period of time; collecting data on the introduction of new products and/ or processes on the market in a certain amount of time; cost reductions made by firms in a certain time period. Those indicators show directly how innovative firms are. My OLS estimates support the idea of having patents as a good estimator for the final innovation effect of R&D subsidies since the fact that a firm has released patents in the last 3 years has a significant effect on the R&D intensity (P=0.000).

(21)

References

Almus and Czarnitzki (2003). The effects of Public R&D Subsidies on Firms Innovation Activities: The Case of Eastern Germany. Journal of Busines & Ecnonomic Statistics. Vol. 21, No. 2: 226-236 Arrow (1962).“Economic Welfare and the Allocation of Resources for Invention”, in The Rate and Direction of Inventive Activity. Princeton University Press: Princeton 1962: 609-625

Bryson, Dorsett, and Purdon (2002): The Use of Propensity Score Matching in the Evaluation of Labour Market Policies. Working Paper No. 4,Department for Work and Pensions.

Busom (1999). An empirical evaluation of the effects of R&D subsidies. Burch center working paper series. No. B99-05

Caliendo and Kopeinig (2005). Some pratical Guidance for the Implementation of Propensity Score Matching. IZA Discussion Paper Series. IZA DP No. 1588: 5-7

Czarnitzki, Ebersberger and Fier(2007). The relationship between R&D collaboration, subsidies and R&D performance: empirical evidence from Finland and Germany. Journal of applied

econometrics. 22.7: 1347-1366

David, Hall and Toole (2000). Is public R&D a complement or substitute for private R&D? A review of the econometric evidence. Research Policy 29: 497-529

Duguet (2004). Are R&D subsidies a substitute or a complement to privately funded R&D? Revue d'économie politique 2/ 2004. Vol. 114 : 245-274

Eurostat and OECD (1997). Proposed guidelines for collecting and interpreting technological innovation data, Paris. OSLO–Manual

González, Jaumandreu and Pazó (2005). Barriers to Innovation and Subsidy Effectiveness. The RAND Journal of Economics. Vol.36, No. 4: 930-950

González, Pazó (2008). Do public subsidies stimulate private R&D spending? Research Policy. 37.3: 371-389

Goolsbee (1998). Does government R&D policy mainly benefit scientists and engineers? No. w6532. National bureau of economic research, 1998

(22)

Heckman, Lalonde and Smith (1999). The Economics and Econo-

metrics of Active Labor Market Programs. Handbook of Labor Economics. Vol.III, ed. by O. Ashenfelter, and D. Card: 1865-2097

Hussinger (2003). R&D and subsidies at the firm level: an application of parametric and semi-parametric two-step selection models. ZEW Discussion Paper. No. 03-63

Klette and Moen (1997), “R&D Investment Responses to R&D subsidies: A Theoretical Analysis and Microeconomic Study”. mimeo 1997

Lach (2002). Do R&D Subsidies Stimulate or Displace Private R&D? Evidence from Israel. The Journal of Industrial Economics. Vol.50, No. 4: 369-390

Lööf and Heshmati (2005). The impact of public funds on private R&D investment: new evidence from a firm level innovation study. MTT Discussion Papers. 3

Nelson, Richard (1959). The Simple Economics of basic Scientific Research. Journal of Political Economy 1959: 297-306.