Cost-Benefit Approach to
Public Support of Private
R&D Activity
Bettina Peters
Centre for European Economic Research (ZEW)
[email protected]
DIMETIC Doctoral European Summer School
Pecs, July 14, 2010
Part II:
Econometrics of Evaluation of Public
Funding Programmes
Motivation
■ Public support of private R&D activity is not without cost either:
crowding-out may occur!
● Once subsidies are available, companies have an incentive to apply for any
project (even for the ones which are also privately profitable) as subsidy
comes at marginal cost equal to zero.
● Subsidies may not only stimulate the projects with high social return.
● In the worst case (total crowding out), private funding is simply replaced with
public funding.
growing literature about evaluation of R&D programmes
The Evaluation Problem
■ The aim of quantitative methods of evaluation is the measurement of
effects generated by policy interventions on certain target variables;
■ We are interested in the causal effect of a treatment 1 relative to
another treatment 0 on the outcome variable Y.
■ In case of public R&D support:
● What is the effect of an R&D subsidy on the subsidized firm‘s R&D
expenses (input)? or
● Or the impact on other variables like patent applications, firm growth,
employment etc. (output)
Different Effects of R&D subsidies:
Self-assessment by companies
(259 subsidized German companies in 2001)
Project implementation became possible
Project start accelerated
Project duration reduced
Project scope extended
Increased technological level
Led to patent application
0% 10% 20% 30% 40% 50% 60% 70%
Source: Czarnitzki et al. (2001)
The Evaluation Problem
■ In most cases, we are interested in the average „treatment effect on the
treated“ (TT)
■ TT: the difference between the actual observed value of the subsidzed
firms and the counterfactual situation:
„Which average value of R&D expenditure would the treated firms
have shown if they had not been treated“
TT  E Y T  Y C | S  1  E Y T | S  1  E Y C | S  1
S:= Status of group, 1 = Treatment group; 0 = Non-treated firms
YT = outcome in case of treatment;
YC = outcome of the treated firm in the case it would not have received the subsidy
(counterfactual situation)
The Evaluation Problem
■ Actual outcome E(YT|S = 1) can be estimated by the sample mean of Y
in the group of treated (subsidized) firms
■ Problem: The counterfactual situation E(YC|S = 1) is never observable
and has to be estimated!
 How to do estimate the counterfactual?
The Evaluation Problem
■ Naive estimator for ATT: Use the average R&D expenditure of nonsubsidized firms assuming that
E Y C | S  1  E Y C | S  0 
■ Assumption is justified in an experiment where subsidies are given
randomly to firms.
■ In real life, however, it is likely that funded firms are typically not a
random sample, but are the result of an underlying selection process
■ Subsidized firms differ from non-subsidized firms
■ It is likely that the Subsidized firms differ from non-subsidized firms and
that the subsidized companies would have spent more on R&D than the
non-subsidized companies even without the subsidy program.
The Evaluation Problem
■ Policy makers want to maximize the probability of success and thus try
to cherry-pick firms with considerable R&D expertise,
● i.e. firms with high high R&D in the past, professional R&D management,
good success with their other R&D projects or experienced in applying for
public funding will be preferably selected.
■ Selection bias in the estimation of the treatment effect.
● We cannot use a random sample of non-treated without any adjustment.
■ As the highest expected success is correlated with current R&D
spending, subsidy becomes an endogenous variable (depending on the
firms characteristics).
■ Solution in non-experimental settings: Microeconometric evaluation
methods (surveys of Heckman et al., 1999; Blundell and Costa-Dias,
2000, 2002).
Microeconometric Evaluation Methods
■ Before-after comparison [panel data]
■ Difference-in-difference estimator (DiD) [panel data]
■ Instrumental Variables estimator (IV) [cross-sectional data]
■ Selection models [cross-sectional data]
■ Matching methods [cross-sectional data]
■ Mixed Method: Conditional difference-in-difference combines the DiD
estimator and matching methods [panel data]
Before-After Comparison
■ Suppose firm i got funding in period t, and we observe R&D expenses in
t and t-1.
■ ATT could be estimated based on the average difference of R&D of
treated firms in t (Yit) and the R&D of the same firm in the previous
period where it did NOT receive a treatment: Yi,t-1.
■ Requires panel data
■ Allows to control for individual fixed effects, but not for macroeconomic
shocks
Difference-in-Difference Estimator
The DiD estimator is based on a „before-and-after“ comparison of
subsidized firms and a non-subsidized control group.
TTDiD  Yt T ,(1)  Yt C ,(1)   Yt C ,(0)  Yt C ,(0) 
1
Advantage:
0
1
0
- no functional form for outcome equation required
- not even a regressor is needed
- controls for common macroeconomic trends
- controls for constant individual-specific unobserved effects
NOTE: when covariates should be included, one can estimate an OLS model in first
differences. (but: functional form assumption necessary!)
Disadvantage: - strategic behavior of firms to enter programs would lead to
biased estimates („Ashenfelter‘s dip“)
- panel data required; including observations BEFORE AND
AFTER (or WHILE) treatment
- biased if reaction to macroeconomic changes differs between
groups (1) and (0)
- problem to construct data if R&D subsidies show high persistence
Instrumental Variables (IV) Estimators
■ Suppose y = b0 + b1 * x1 + u
■ We think that x1 is endogenous, i.e. COV(x,u)!=0.
● e.g. wage equation:
• wage may depend on education and ability.
• But we only observe x1 = education.
• Then, u = v + b2*ability (where v is a new error term, b2 is the
coefficient of ability)
■ OLS would be inconsistent as it relies on COV(x,u)=0.
■ Suppose we have an instrument „w“, that fullfils two requirements:
● w is uncorrelated with the error term u  COV(w,u)=0, i.e.
(i.e. z should have no partial effect on y once we control for x1)
● and w is correlated with the endogenous variable x, i.e. COV(w,x)!=0.
■ IV estimator:
ˆb   WX 1 Wy
IV
Instrumental Variables (IV) Estimators
■ The recent utilization of IV estimators in context of evaluation goes back
to Imbens/Angrist (1994) and Angrist et al. (1996) who invent the Local
Average Treatment Effect (LATE).
■ IV estimators have the advantage over selection models that one does
not have to model the selection process and to impose distributional
assumptions.
■ Main disadvantage: need of an instrument, whose requirement are more
demanding than those for the exclusion restriction in selection models.
■ Instruments can be
●
●
other variables (external instruments, often hard to find and justify)
lagged values of endogenous variables (requires panel data)
■ In the case of R&D it is very difficult to find valid instruments.
Selection models
■ control function approach
■ Selection models are based on a two step procedure
(based on Heckman‘s work, 1974, 1976, 1979):
● estimate the propensity to get an R&D subsidy for all firms
Si*  Z i  Vi
1, if Si*  0,
Si  
0, otherwise
● estimate outcome equation for participants and non-participants including a
correction for a possible selection mechanism,
Yi1  X i11  U i1 ,
if Si  1
Yi 0  X i 0 0  U i 0 ,
if Si  0
Selection models
■ Under the assumption of joint normality, we can estimate:
  Z i 
E Yi1 | X i1 , Si  1  X i11  1 1
  Z i 
E Yi 0 | X i 0 , Si  0   X i 0  0  0 0
  Z i 
1    Z i 
■ Madalla (1983): ATT is determined by subtracting the estimated R&D
expenditure of publicly funded firms, which they would have conducted
if they had not received public R&D funding, from the expected R&D
expenditure of funded firms. The difference is augmented by the
selection correction
ˆTT
^
 ^
   Z i 
ˆ
ˆ
 X i 1   0   1 1   0 0 

   Z i 


Selection models
■ Advantage:
● Controls for unobserved characteristics (entering the first- and second-step
equation).
● Root-N-consistency
■ Disadvantage:
● Restrictive distributional assumption on the error terms (joint normality).
● An exclusion restriction is needed which is included in the selection
equation but not in the structural equation to identify the treatment effects.
● A fully parametric model for the selection and for the structural equation has
to be defined.
Semiparametric Selection Models
■ Semiparametric estimators: Gallant and Nychka (1987), Cosslett (1991),
Newey (1999), or Robinson's (1988) partial linear model.
■ Semiparametric estimators identify only the slope parameters of the
outcome equation. Intercept in outcome equation is no longer identified,
but required for deriving ATT
■ An additional estimator for the intercept is needed to identify the
treatment effects, e.g. Heckman (1990),Andrews and Schafgans (1998).
■ See Hussinger (2008) for applications of such estimators for the
evaluation of innovation policy.
Matching
■ Ex post mimic an experiment by constructing a suitable control group by
matching treated and non-treated firms
■ Selected control group is as similar as possible to treatment group in
terms of observable characteristics.
■ Matching is a nonparametric method to identify the treatment effect
YiT  g T  X i   U iT if Si  1,
Yi C  g C  X i   U iC if Si  0.
Matching
■ A1: Conditional independence assumption (CIA) (Rubin 1974, 1977): All
the relevant differences between the treated and non-treated firms are
captured in their observable characteristics
E Y C | S  1, X  x   E Y C | S  0, X  x 
=> For each treated firm, search for twins in the „potential control group“
having the same characteristics, X, as the subsidized firms.
■ A2: We observe treated and non-treated firms with the same
characteristics (common support)
■ Under these assumptions, the ATT can be calculated as:
TT  E Y T | S  1, X  x   E Y C | S  0, X  x 
Matching
Yi T   wijY jC
■ Treatment effect for firm i:
j
■ Two common matching estimators:
● Nearest Neighbor: wij=1 for the most similar firm, zero otherwise
=> only one control observation is used
● Kernel-based:
entire control group is used for each treated firm,
weights wij are determined by a kernel that
downweights distant observations from Xi.
wij
w12
w11
xi - xj
Kernel-Based Matching
■ Weights are the kernel density at Xj - Xi (rescaled that they sum up to 1)
w ij 

K h 1  X j  X i 
 K h  X
1
j

 Xi 

j
■ Often the Gaussian kernel or the Epanechnikov kernel is used,
■ Calculation of counterfactual requires kernel-regression (e.g. NadarayaWatson estimator)

m  X i   min  Y  m  K h 1  X j  X i 
m
C
j
2

j
locally weighted average of the entire control group (for
each treated firm)
K h 1 X  X
 
 Kh X
  
j
j
1
j
j

Y  w Y
 X 
i
j
i
ij
j
j
Kernel-Based Matching
■ Bandwith h may be chosen according to Silverman‘s rule of thumb:
0.9 An 1/ 5 if k  1
h
1/ 5 2
k
0.9
n
 if k  1
 
● with k : number of arguments in the matching function
■ If you want to include more than a single X in the matching function, you
can use the Mahalanobis distance
MDij   X j  X i  1  X j  X i 
'
Propensity Score
■ Usually X contains many variables which make it almost impossible to
find control observations that exactly fit those characteristics of the
subsidized firm.
■ Rosenbaum and Rubin (1983) showed that it is possible to reduce X to
a single index - the propensity score P - and match on this index.
■ It is possible to impose further restrictions on the control group, e.g.
that a control observations belongs to the same industry or same region
etc.
A NN Matching Procedure
1. Specify and estimate probit model to obtain propensity scores
2. Restrict sample to common support:
■ Delete all observations on treated firms with propensity scores larger than the
maximum and smaller than the minimum in the potential control group.
■ Do the same step for other variables that are possibly used in addition to the
propensity score as matching argument.
3. Choose one observation from sub sample of treated firms and delete it from
that pool
4. Calculate the Mahalanobis distance between this treated firm and all nonsubsidized firms in order to find the most similar control observation.
MDij   Z j  Zi  1  Z j  Zi 
'
■ Z contains the matching arguments (propensity score and/or additional variables
such as e.g industry or size classes)
■ Ω is the empirical covariance matrix of the matching arguments based on the sample
of potential controls
A NN Matching Procedure
5. Select observation with minimum distance from potential control group as twin
for the treated firm
■ NN matching with replacement: selected controls are not deleted from the set of
potential control group so that they can be used again
■ NN matching without replacement: selected controls are deleted from the set of
potential control group so that they cannot be used again
6. Repeat steps 3 to 5 for all observations on subsidized firms
7. The average effect on the treated = mean difference of matched samples:
ˆTT 

T
C
ˆ
Y

Y
i i 
T  i
n  i
1
■ With YC_hat being the counterfactual for firm i and nT is the sample size of treated
firms.
8. Sampling with replacement  ordinary t-statistic on mean differences is biased
(neglects appearance of repeated observations)  correct standard errors:
Lechner (2001)  estimator for an asymptotic approximation of the standard
errors
Matching in Stata
■ Psmatch2.ado
■ Software and documentation from Barbara Sianesi and Edwin Leuven,
IFS London
http://www.ifs.org.uk/publications.php?publication_id=2684
http://ideas.repec.org/c/boc/bocode/s432001.html
Disadvantages of Matching
■ It only allows controlling for observed heterogeneity among treated and
untreated firms (in observable cahracteristics in X)
■ „Common support“ is necessary, that is, the range of the propensity
score of the control group must cover the treatment group.
● If the common support is rather small in your data, matching is not
applicable
Mixed method:
Conditional Difference-in-Difference
■ Conditional difference-in-difference (DiD) method for repeated crosssections, which combines ordinary DiD estimation with matching
■ The Conditional DiD estimator consists of matching firms i and j with the
same observable characteristics X_i,t0= X_j,t0 where i receives
treatment in t1 but not in t0 and j is a non-treated firm in both periods.

CDiD
TT

 Y Y
T
i ,t1
C
i ,t0
  Y
C
j ,t1
Y
C
j ,t 0

■ Heckman et al. (1998) show that CDiD based on non-parametric
matching proved to be a very effective tool in controlling for both
selection on observables and unobservables.
Microeconometric Evaluation Methods
■ Before-after comparison [panel data]
■ Difference-in-difference estimator (DiD) [panel data]
■ Instrumental Variables estimator (IV) [cross-sectional data]
■ Selection models [cross-sectional data]
■ Matching methods [cross-sectional data]
■ Mixed Method: Conditional difference-in-difference combines the DiD
estimator and matching methods [panel data]
Which method to use?
■ The econometric method that you can apply heavily depends on the
data you have:
● Panel or cross-section?
● Is the treatment variable a binary indicator (yes/no) or is it a continuous
treatment variable?
● Do I have candidates for instrumental variables?
■ Do I want to make functional form assumptions of my R&D investment
equation?
■ Do I want to specify a structural model or simultaneous equation
system?
Empirical Studies
■
Busom (2000), 154 obs., Spanish manufacturing, parametric selection model
■
Wallsten (2000), 479 obs., US SBIR program, simultaneous
equations model, 3SLS (incl. amount of funding)
■
Czarnitzki (2001), 640 obs., Eastern German manufacturing, NN-Matching
■
Czarnitzki/Fier (2002), 1,084 obs., German service sector, NN-Matching
■
Fier (2002), 3,136 obs., German manufacturing (specific program), NN-Matching
■
Lach (2002), 134 obs. Israeli manufacturing, DiD and dynamic panel models
■
Almus/Czarnitzki (2003), 925 obs., Eastern German mf., NN-matching
■
Gonzales et al. (2006), 2.214 obs. Spanish manufacturing, simultaneous
equations model with thresholds:
■
Hussinger (2008), 3744 obs., German manufacturing sector 1992-2000, parametric and
semiparametric selection models
■
Schmidt and Aerts (2008), Germany and Flanders, CIS3+4, NN matching and CDID
■ Surveys: David et al. (2000; survey on crowding-out effects), Klette et
al. (2000, including output analyzes like firm growth, firm value, patents
etc.), Parsons and Phillips (2007), Aerts et al. (2007)
Example for Effect of R&D subsidies on R&D
Expenditure Using Matching Estimators
Schmidt and Aerts (2008), Two for the price of one? Additionality effects of R&D
subsidies: A comparison between Flanders and Germany, Research Policy 37
(5), 806-822
■ Data: German and Flemish Community Innovation Surveys (3 and 4)
■ 2 methods:
● Matching estimator and
● conditional DiD
Mean Comparison Before Matching
Mean Comparison Before Matching
Probit Estimations and Marginal Effects
Mean Comparison After Matching
Mean Comparison After Matching
Average Treatment Effects of the
Treated Companies
To sum up: Does public funding stimulate or
crowd out private R&D expenditure?
■ Nearly all empirical studies reject the hypothesis of a total crowding out
(i.e. no change in total private R&D expenditure due to public funding).
● Exception: Wallsten (2000) for US SBIR program
■ Hypothesis of partial crowding out is also often rejected.
● David et al. (2000): At the macro level, only 2 out of 14 studies yield a
substitute relationship of public and private R&D investment. At the firm
level: 9 out of 19.
■ Czarnitzki et al. (2002): average multiplier effect of 1 which can be
higher for specific groups
To sum up: Does public funding stimulate or
crowd out private R&D expenditure?
■ Crowding in effects: Public R&D subsidies stimulates net R&D
expenditure (total R&D exp. minus subsidy):
● Gonzales et al. (2006): multiplier effect for Spanish firms in 1990-1999
slightly above 1
● Fier et al. (2004): multiplier effect of 1,14 for German firms in 1990-2000
(varies according to technology fields)
● Hussinger and Czarnitzki (2004): multiplier effect of 1.44
● Parsons and Phillips (2007): average multiplier effect of 1.29 for surveyed
studies
■ Large variation in estimated multiplier effect, not surprising because
funding schemes are different and have to be taken into account.
Extensions
■ Heterogeneous treatments
■ Effects on innovation output
■ Effects on innovation behaviour
Heterogeneous Treatments
■ So far, simply binary indicator (funded yes/no)
■ Heterogeneous Treatments, e.g.
● Countinuous treatment
● Categorial treatment
■ Countinuous treatment
● Hirano and Imbens (2005)
● different subsidies levels
● generalized propensity score (GPS) method for the estimation of so called
dose-response functions.
■ Categorial treatment
● Imbens (2000), Gerfin and Lechner (2002):
● divide treated firms in different groups, e.g. low subsidy and high subsidy
● distinguish between different policy programs.
Effects on Innovation Output
(Output Additionality)
■ Subsidies may just increase wages of R&D employees but not the
number of R&D personnel. If an increase in wages does not go along
with higher research productivity, subsidies are likely to result in higher
innovation input, but not necessarily in innovation output.
■ Subsidized projects may be associated with higher risk than privately
financed projects. If failure rates are higher, subsidies are likely to result
in higher R&D investment, but not necessarily in innovation output.
● Czarnitzki and Hussinger (2004) and Czarnitzki and Licht (2006) add patent
equation to the input model. Both purely private R&D and publicly funded
R&D increase patenting output. Subsidized R&D is a little less productive,
though.
Effects on Innovation Behaviour
(Behavioural Additionality)
■ Example: Current practice in Europe is to support research consortia
(firms+firms / firms + scientific institutions) rather than giving subsidies
to individual firms.
■ Czarnitzki, Ebersberger and Fier (2007) apply Gerfin/Lechner
methodology to investigate effects of subsidies vs. R&D collaborations
in Germany and Finland
● R&D collaboration achieves R&D input (and output) more than subsidies to
individual firms
● AND there is room for fostering collaboration especially in Germany
(“Treatment effect on the untreated”).
Table 6:
Matching Results for Germany: Average Treatment Effects E( m,l)
Dependent variable: R&D Intensity (R&D expenditures/Sales * 100)
Actual state (m)
None
None
Counterfactual
state (l)
Collaboration
Public funding
Both
1.591
Collaboration
1)
0.775**
(0.035)
4)
-0.096
(0.325)
7)
-0.276
(0.506)
10)
-2.420*** 11) -2.143***
(0.842)
(0.792)
2.224
8)
0.335
(0.480)
Public funding
Both
2)
0.278
(0.424)
3)
2.533***
(0.056)
5)
0.103
(0.596)
6)
2.195***
(0.607)
9)
1.903***
(0.654)
2.497
12) -1.608**
(0.820)
4.821
Further challenges in research
■ Output effect mainly measured in terms of patents (patents are an
indicator of inventions, not necessarily of innovations)
● Alternative innovation out: share of sales with new products (Hussinger
2008)
■ Empirical: Does collaborative R&D funding result in collusion in product
market?
● Overall welfare effect might be negative
■ Specifities of policy schemes are not exploited in current research.
● Ideally, policy makers would like to know if a certain program design is more
likely to prevent crowding-out effects than another.
■ Selection equation is a reduced form estimation. Decision of the firm to
apply and decision of the government to support a firm are not
separately accounted for.
● Solution: Structural models (see work by Otto Toivanen)
More Literature
■ NBER Summer Course in Econometrics by Guido Imbens and Jeff
Wooldridge
● http://www.nber.org/minicourse3.html
● Includes videos of lectures and extensive lecture notes
■ Survey by Imbens and Wooldridge (2009)
● http://www.economics.harvard.edu/faculty/imbens/files/recent_development
s_econometrics.pdf
● published in Journal of Economic Literature
Part III:
Simple Cost-Benefit Approach to
Public Support of Private
R&D Activity
State Aid for R&D&I
Community Regulation Dec. 30, 2006
■ „State aid for R&D&I shall be compatible if the aid can be expected
to lead to additional R&D&I and if the distortion of competition is not
considered to be contrary to the common interest, which the
Commission equates for the purposes of this framework with
economic efficiency“
■ „To establish rules ensuring that aid measures achieve this
objective, it is, first of all, necessary to identify the market failures
hampering R&D&I“
■ “Negative effects of the aid to R&D&I must be limited so that the
overall balance is positive”.
Cost-Benefit Analysis
■ Empirical evidence that social returns to R&D exceed private returns
identifies market failure and thus provide a central argument in favour of
direct or indirect public support of private R&D activities.
■ But: only necessary but sufficient condition
■ Public R&D programmes are always associated with costs which go
beyond the pure size of the subsidy
■ Cost-Benefit-Analysis:
● Evaluation of taxed-based R&D support (R&D tax credits) in Australia
(Lattimore 1997), Netherlands (Cornet 2001a,b) and Canada (Parsons and
Phillips 2007)
● Evaluation of direct project-based R&D support in Germany (Peters,
Kladroba, Licht, Crass 2009)
Cost-Benefit Analysis
Basic idea:
■ Government supports R&D&I activities of firms in period t=0 (size of the
support: P=1 €)
■ No returns to R&D in funding period (t=0)
■ Returns Rt accruing from R&D from period t=1 onwards
■ Compare net present value of benefits and costs
● A project is beneficial if C0 is larger than 0 or equivalently benefit/cost-ratio
is larger than 1
C0   K0  B0
Benefits
■ Returns:
● in period t=1: R
● in period t=2: R*(1-d), where d is the depreciation rate on knowledge
● …
■ What is R?
● Returns R to R&D are equal to the actual change of private R&D
expenditure times the average social rate of return s.
● Change of private R&D expenditure depends on size of public support P
and multiplier/crowding out effects m
● Further account for the fact that a proportion λ of the subsidies may just use
to increase wages of R&D employees but not to increase the amount of
research that is undertaken.
Benefits
■ Returns are discounted with discount factor i, consisting of
● the time preference rate r (reflecting e.g. the interest rate of risk-free
investment) and
● the risk premium π (additional return a firm requires to invest in risky R&D
projects)
■ Present value of the benefits of subsidizing private R&D having a finite
time-horizon of T:
T
(1  r   )T  1  d 
B0   s  m  (1   )  P  
  P
T
(1  r   )  (d  r   )


■ … having an infinite time-horizon:
B0   s  m  (1   )  P  
1
  P
(d  r   )
Benefits
■ Alternative assumptions about multiplier effects m based on
econometric evaluation studies:
●
●
●
●
●
●
0
0.6
0.9
1.0
1.15
1.3
● 2
(total crowding out)
(strong crowding out)
(weak crowding out)
(weak crowding in, preferred conservative estimate)
(medium crowding in; average estimated effect reported in the survey
by Parson & Philips 2007)
(strong crowding in)
■ Social rate of return s based on spillover literature:
● Assumption: additionally publicly funded R&D yields the same average
social return
● Preferred assumption: s=0.5 (alternatives: 0.15 / 0.3 / 0.7 and 1.0)
Benefits
■ Wage elasticity of labour λ:
● Goolsbee (1998): based on data of 17,700 US scientists and engineers from
the years 1968 to 1994 he estimates that an increase of public R&D funding
by 11% increase wages on average by 3.3% (wage elasticity varies
between 2 and 6 % depending on educational background).
● Given that 2/3 of R&D expenditure is for labour, they estimated the actual
increase in research to be roughly 23% lower.
● Marey and Borghans (2000): 20-30% of increase in R&D expenditures due
to introduction of tax credits is related to higher wages
● Effect will depend on labour market specifities
● Preferred assumption: 10%; alternative assumptions: 5%, 20% and 30%
Benefits
■ Depreciation rate d:
● d=15% (alternatively: 10%, 20%)
■ Time preference rate r:
● r=3.5% (alternatively: 5%)
■ Risk premium π:
● π =3% (alternatively: 5%)
■ Time horizon:
● T=15 (alternatively: 5, 10, 20 years and infinite horizon)
Costs
■ Different types of costs:
●
●
●
●
Direct programme costs (P) in period t=0 (size of subsidy or forgone taxes)
Administrative costs of government
Administrative costs of firms
Tax funding of subsidies induce a distortion of resource allocation welfare
loss (marginal excess burden)
● Forgone returns of an alternative investment
■ Present value of costs
K 0  1  tx   1  cs   P  cu  m  P 
●
●
●
●
   m  1  1     P
d  r   
cs = public administrative costs;
cu = administrative costs of the firm;
tx = macroeconomic costs of tax financing;
β = return to the alternative investment
Costs
■ Administrative costs
● Till now scarcely evaluated: Gunz et. al. (1997) and Parsons and Phillips
(2007) for Canada
● Administrative costs of firms
• Administrative costs of the firms varies with the policy measure
• They are expected to be much lower with R&D tax credits than with
R&D project funding (require less paperwork and entail fewer layers of
bureaucracy)
• Proportion of the administrative costs decreases with absolute project
size resp. absolute altitude of the tax abatement
– For project based funding: 3-25% of support received (average 8%)
– For fiscal funding: 15% (10%, 5%) of the tax abatement if tax abatement
(<100,000 $, 100,000-500,000 $, >500,000 $)
• Basic specification: cU=8% (alternative assumptions: 5%, 10%, 20%)
Costs
● Administrative costs to government:
– 1.7% related to the whole taxes foregone in case of tax credits;
– 3-8% in case of project based funding
• Basic specification: cS=3% (alternative assumptions: 2%, 5%, 10%)
■ Macroeconomic costs of tax financing
● Lattimore (1997) estimated welfare losses due to distortive effects of tax
financing of public funding: 15-50% of direct program costs
● Distortive effect depends on type of taxes raised
● Parsons and Phillips (2007) estimated a distortive effect of 27%.
• Basic specification: tx=30% (alternative assumptions: 15%, 50%)
■ Return to the alternative investment:
• Basic specification: β=5%
A Simple Cost-Benefit Approach to
Public Support of Private R&D Activity
Benefit-to-Cost-ratio
Multiplier
0,0
0,6
0,9
1,0
1,15
1,3
2,0
0,15
0,09
0,35
0,46
0,50
0,55
0,59
0,77
Social rate of return
0,3
0,5
0,7
0,09
0,09
0,09
0,63
1,00
1,37
0,86
1,38
1,90
0,93
1,50
2,07
1,02
2,30
1,66
1,12
1,82
2,52
1,49
2,44
3,40
1,0
0,09
1,93
2,69
2,92
3,25
3,57
4,83
Note: Assumptions on parameters: P=1, d=0,15,r=0.035,
TT=0.035, cU=0.08, cS=0.03, tx=0,3,λ=0.1 and T=15.
Source: Peters et al. (2009)
Using the preferred parameter estimates, benefits of public R&D
subsidies would exceed costs by roughly 1.66
Multiplier, Social Benefits and the Cost-BenefitRelationship of Public R&D Funding
area of probable combinations of
multiplier effects and social benefit rates
3
Multiplikator
multiplier
2,5
2
macroeconomic costs =
macroeconomic benefits
1,5
1
0,5
0
0
0,5
1
Soziale
Erträge
social
rate of
return
1,5
Effect of Time Preference Rate, Risk Premium
and Depreciation Rate
Time preference
rate r
0,05
0,035
BenefitCostRatio
1,66
1,57
Risk premium

Depreciation rate d
0,03
0,05
0,15
0,1
0,2
1,66
1,54
1,66
2,03
1,39
Additional parameter assumptions: m=1.15, s=0,5, P=1,cU=0.08, cS=0.03, tx=0.3, λ=0.1,
β=0.05 and T=15.
Source: Peters et al. (2009)
Effect of Tax Distortion and
Administrative Costs
Tax distortion tx
0,30
0,15
0,50
Administrative
costs of firms c
0,08 0,10 0,20
1,66
1,86
1,46
1,66
U
BenefitCost-Ratio
1,63
1,52
Administrative costs of
government c
0,03 0,02 0,05 0,10
S
1,66
1,68
1,63
Additional parameter assumptions: m=1.15, s=0,5, P=1,r=0.035, π=0.03, d=0.15, tx=0.3,
λ=0.1, β=0.05 and T=15.
Source: Peters et al. (2009)
1,56
Effect of Wage Elasticity and Time Horizon
0,10
BenefitCost-Ratio
1,66
Wage elasticity
0,05
0,20
1,71
1,56

0,30
15
1,45
1,66
Time horizon
5
10
1,18
1,54
Additional parameter assumptions: m=1.15, s=0,5, P=1,cU=0.08, cS=0.03, r=0.035,
π=0.03, d=0.15,β=0.05 and tx=0.3.
Source: Peters et al. (2009)
20
1,70
Summary
■ Using the preferred parameter estimates, benefits of public R&D
subsidies would exceed costs by roughly 1.66.
■ Positive effects for a broad range of parameter values.
■ Overall effect of public R&D funding crucially depend on the amount of
social returns to R&D and multiplier effects.
■ Even in case of low social returns to R&D public funding might be
beneficial, the likelihood increases with increasing multiplier effects.
■ Crowding out effects do not necessarily imply a welfare loss (v.v.)
● E.g. strong crowding out effects (m=0.6) could be compensated by an
average social rate of return of 0.5.
■ Other parameters are less important.
● Only a modest impact of time preference rate, risk premium and
administrative costs.
● Moderate impact of depreciation rate, tax distortion and time horizon
Limitations
■ Ideally, policy makers would like to know which program design is
presumably the most efficient.
■ Program-specific cost-benefit analysis would require program-specific
estimates of model parameters (multipliers, social rates of return, …)
●
not yet available
■ It is argued that there is presumably a trade-off:
● Social benefits are expected to be higher for tax credits whereas multiplier
effects are expected to be higher for public subsidies.
Back-up slide
Aid Intensities within EU State Aid Rules
Share of Public Funding in Total Project Costs
Small
enterprise
Mediumsized
enterprise
Large
enterprise
Fundamental research
100%
100%
100%
Industrial research
65%
60%
50%
80%
75%
65%
40%
35%
25%
55%
50%
40%
- collaboration between undertakings; for large
undertakings: crossborder or with at least one
SME
- collaboration of an undertaking with a public
research organisation
- dissemination of results
Experimental development
- collaboration between undertakings; for large
undertakings, with cross-border or at least one
SME
- collaboration of an undertaking with a public
research organisation