1

Identifying Technology Spilloversand Product Market Rivalry

Nick Bloom (Stanford & CEP) Mark Schankerman (LSE and CEPR)

John Van Reenan (LSE and CEP)

Conference on Productivity Growth:Causes and Consequences

San Francisco Federal Reserve Bank

18-19 November 2005

2

IntroductionTwo main types of R&D “spillover” effects

1. technological spillovers (Griliches, 1992; Keller,2004)2. product market rivalry effect (I.O. models of R&D)

• Direct business stealing effect on firm profits• Strategic effect on R&D decisions

Why try to identify and quantify their separate effects?• Current estimates of technological spillovers may be

contaminated by product market rivalry effects• In order to assess the net effect of spillovers: Is there

over- or under- investment in R&D?• Useful for examining different technology policy analysis

(e.g. R&D subsidies for smaller vs. larger firms)

3

Empirical Identification Scheme• We use separate measures of technological closeness

(through multiple patent classes) and product market closeness (through multiple industry classes)(Bransetter and Sakakibara, 2002)

• We use multiple outcome variables (market value, patents and R&D). Also look at productivity– Identification of technological spillovers comes from

the R&D, patents and value equations. – Identification of strategic rivalry effect comes from

value eq (existence) and R&D equation (direction – complements or substitutes).

(Griliches, Hall and Pakes, 1991)

4

Findings

• Empirical evidence of both technology spillovers and product market rivalry (controlling for unobserved heterogeneity is important)

• Technology spillover effects dominate – social return to R&D over three times the private return

• See this in pooled sample and in analysis for 3 high tech sectors

• Simulation of R&D policy suggests that medium sized firms generate less spillovers than large firms (EU policies)

5

Structure

1. Intro2. Analytical Framework3. Data4. Econometrics5. Results6. Extensions – Industry heterogeneity,

policy simulations7. Conclusions

6

2. Analytical FrameworkTwo stage game. R&D (r) choice at stage 1. Knowledge, k (we will proxy by

patents) then revealed as a function of firms’ R&D. Firms choose price/quantity (x) at Stage 2.

Three firms: 0, τ and m. Firms 0 and m compete in the same product market. Firms 0 and τ operate in same technology area. Can generalise to many firms interacting in both product and technology markets.

Stage 2: Short run competition

• Profit of firm 0 depends on short run decision variable by firm 0 (x0 ), her product market rival (xm) and on the knowledge stock of firm 0 (k0). We assume Nash Equilibrium but make no assumption on what x is (e.g. Cournot or Bertrand).

• The best responses on x0 and xm yield a reduced-form profit function for firm 0 (and firm m) that depends on k0 and km: Π0(k0,km) = π (k0,x*0,x*m)

• Firm 0’s profit increases in k0 and declines in km . The latter is the direct business stealing effect.

• We allow for either strategic substitutability or strategic complementarity in R&D

7

Stage 1: Production of knowledge (k)

• k0 is produced with its own (r0) and firm τ’s R&D (rτ): k0 = Φ(r0,rτ). k0 is increasing in both arguments. rτ may increase, reduce or leave unchanged the marginal product of r0 .

• We analyse how rτ and rm affect the best response R&D, r0 , knowledge (patent) k0, and the value of the firm V0.

• Look at this non-tournament model. Compare to a tournament (patent race) model in Appendix A. The main empirical predictions are similar.

8

Some basic Predictions (Table 1)

• With any SIPM (strategic interaction in product market) market value (V) falls with rival R&D (rm)

• With any technology spillovers market value rises with technological neighbor’s R&D (rτ)

• Knowledge, k0 (patents) increases with technological neighbor’s R&D rτ if’f there are tech spillovers

• Knowledge is unaffected by rm

• If strategic interaction in product market own R&D (r0) rises with rival R&D (rm) if strategic comps and falls with rm if strategic subs

9

Comparative static prediction

Empirical counterpart

No Technological Spillovers

No Technological Spillovers

Some Technological Spillovers

Some Technological Spillovers

Strategic complements

Strategic Substitutes

Strategic complements

Strategic Substitutes

∂V0/∂rτ Market value with SPILLTECH

Zero Zero Positive Positive

∂V0/∂rm Market value with SPILLSIC

Negative Negative Negative Negative

∂k0/∂rτ Patents with SPILLTECH

Zero Zero Positive Positive

∂k0/∂rm Patents with SPILLSIC

Zero Zero Zero Zero

∂r0/∂rτ R&D with SPILLTECH

Zero Zero Ambiguous Ambiguous

∂r0/∂rm R&D with SPILLSIC

Positive Negative Positive Negative

Table 1. R&D Spillovers and Strategic Rivalry

10

Note: Ambiguity of effect of tech neighbor’s R&D on own R&D

• ∂r*0/∂rτ

• Depends partly on sign impact of neighbours R&D on marginal product of own R&D (∂2Φ/ ∂ r0∂

rτ)

– Could be positive if own marginal productivity of R&D increasing in neighbour’s R&D

– Could be negative if “fishing out” of ideas (own marginal productivity of R&D decreasing in neighbour’s R&D)

• Sign also depends on diminishing returns to knowledge production. Will be negative if these are very strong

11

3. Data

• Compustat (all listed US firms) for R&D, Tobin’s Q, capital, labor and distribution of sales by SIC’s (Compustat Line of Business files)– Sample period: 1980-2001– Sample period for sales by 597 4-digit SIC classes: 1993-2001.

Average number of classes per firm is 5.2 (range is 1 to 36)

• US Patents and Trademarks Office (USPTO) for patent counts, patent citations and distribution of patents by tech classes (Jaffe and Trajtenberg, 2002)– We keep firms with at least one patents between 1968 and1999.– Patents assigned to 426 technology classes. Average number

of patent classes 33 (range is 1 to 320)

12

Technology Spillovers

• Following Jaffe (1986) we compute technological closeness by uncentred correlation between all firm pairings of patent distributions

• Define Ti = (Ti1, Ti2 ,, ……, Ti426) ; where Tik is the % of firm i’s total patents in technology class k (k = 1,..,426) averaged 1968-1999. e.g. if all patents in class 1, Ti= (1,0,0,…..0)

• TECHi,j = (Ti T’j)/[(Ti Ti’)1/2(Tj T’j)1/2]; ranges between 0 and 1 for any firm pair i and j

• Technology spillovers are defined as:SPILLTECHit = Σj,j≠iTECHi,jGjt

where Gjt is the R&D stock of firm j at time t (Gjt = Rjt + (1-δ) Gjt-1)

13

Product Market Spillovers

Analogous construction of product market “closeness”

Define Si = (Si1, Si2 ,, ……, Si597) where Sik is the % of firm i’s total sales in 4 digit industry k

(k = 1,…,597)

SICi,j = (Si S’j)/[(Si Si’)1/2(Sj S’j)1/2]

Product market “spillovers” from R&D are

SPILLSICit = Σj,j≠i SICi,j Gjt

14

Identification of product market and technological spillovers

• How distinct are TECH and SIC?• TECH,SIC Correlation is only 0.47 (see

figure 1)• SPILLTECH, SPILLSIC correlation is 0.42

in cross section and 0.17 in growth rates (latter relevant for empirics which control for fixed effects)

• Examples

15

Figure 1: Correlation between SIC and TEC across all firm pairs

Technological closeness

Product market closeness

16

Examples (High TECH, low SIC): Computer and chip makers

Correlation IBM Apple Motorola Intel

IBM SICTECH

11

0.320.64

0.010.47

0.010.76

Apple SICTECH

11

0.020.17

0.010.47

Motorola SICTECH

11

0.350.46

Intel SICTECH

11

IBM, Apple, Motorola and Intel all close in TECH (sample mean .13)But a) IBM close to Apple in product market (.32, computers; sample mean=.05) b) IBM not close to Motorola or Intel in product market (.01)

Table A1

17

Other examples (high SIC, low TEC)

• Gillette corp. and Valance Technologies compete in batteries (SIC=.33, TECH=.01). Gillette owns Duracell but does no R&D in this area (mainly personal care products). Valence Technologies uses a new phosphate technology to improve the performance of standard Lithium ion technology

• High end hard disks. Segway with magnetic technology Philips with holographic technology

• HDTVs

18

4. Econometric Specifications: Generic issues across all 4 equations

• Unobserved heterogeneity (correlated fixed effects)

• Endogeneity (use lagged x’s, but also experiment with IV/GMM approach)

• Dynamics – allow lagged dependent variable• Demand controls – time dummies, industry sales• Newey-West corrected standard errors

ititit uxy '

19

(a) Market Value

it

vit

it A

G

A

V 1lnlnln

Griliches (1981)

vitt

vi

vvvitit

v

itv

itit

vZSPILLSIC

SPILLTECHA

G

A

V

312

11

1

'ln

lnln

6th order series expansion (compare with NLLS)

R&D stock/Fixed assets

Fixed effects(long T)

Time dummies

Tobin’s Q (market valueto fixed assets)

Should be +ve

Should be -ve

20

(b) R&D Equation

ritt

ri

rrritit

r

itr

itr

it

vZSPILLSIC

SPILLTECHRR

312

111

'ln

lnln

R&D expenditures

Test of strategic complementarity (βr2>0) vs. strategic substitutes (βr

2<0)

Compare with Blundell Bond (1998, 2000) GMM-SYS approach

21

(c) Patent Count Equation

p

ittp

ippp

ititp

itp

itititititit

vZSPILLSIC

SPILLTECHDPDPXPE

312

112111

'ln

lnlnexp),|(

• Note: Own R&D stock in Z; D=1 if P>0, D=0, otherwise (MFM) • Allow for overdispersion via Negative Binomial model

• Use Blundell, Griffith and Van Reenen (1999) control for fixed effectswith weakly exogenous variables through pre-sample mean patents (1968-1984)

• Compare with Linear Feedback model estimated by GMM

• Compare with citation weighted patents as dependent variable

+ve

0

22

(d) Production Function

yitt

yi

yyyitit

y

ity

it

vZSPILLSIC

SPILLTECHY

312

11

'ln

lnln

Deflated sales

Own R&D, labour, capital, etc.

• problem of firm specific prices• Allow different coefficients on factor inputs in different industries• Compare with Blundell Bond (1998, 2000) GMM-SYS approach and Olley-Pakes (1995) extension

+ve

0

23

Reflection Problem

Manski (1991) - Identifying endogenous social effect (R&D spillovers)

Main concerns: technological opportunity (supply) and demand shocks that are common to firms.

To address we include:• Parametric determination of who are the “neighbors” (cf. Slade,

2003)• Industry sales variables: constructed using firm’s distribution of

sales across SIC’s• Fixed firm effects• Value function w.r.t. rival R&D robust to this critique (but R&D

equation most problematic)

There remains an issue: Do we identify technology spillovers or supply shocks?

24

5. Main Results

• Tobin’s Q equation

• Patents equation

• R&D equation

• Productivity equation

25

Table 3: Tobin’s QDependent variable:

Ln (V/A)(1) (2) (3) (4)

No individual Effects

Fixed Effects Fixed Effects (drop SPILLSIC)

Fixed Effects (drop

SPILLTEC)

Ln(SPILLTECHt-1)

Technology spillovers

-0.040(0.012)

0.240(0.104)

0.186(0.100)

Ln(SPILLSICt-1)

Product market rivalry

0.038(0.007)

-0.067(0.031)

-0.047(0.031)

Ln(Industry Salest) 0.434(0.068)

0.294(0.044)

0.298(0.044)

0.299(0.044)

Ln(Industry Salest-1) -0.502(0.067)

-0.170(0.045)

-0.176(0.045)

-0.164(0.043)

Note: Sixth-order terms in ln(R&D/Capital) and time dummies also included. NT=10,011

26

Quantification of value eq

• $1 of own R&D associated with $1.18 higher V (cf. Hall, Jaffe, Trajtenberg, 2005, $0.86)

• $1 of SPILLTECH associated with 4.32 cents higher V

• $1 of SPILLTECH is worth 3.6 percent as much as own R&D

• $1 of SPILLSIC associated with 4.36 cents lower V

• Industry sales growth raises value, conditional on R&D variables

27

Table 4: Patent Model (1) (2) (3) (4) Dependent variable: Patent Count

No initial conditions: Static

Initial Conditions: Static

Initial Conditions: Dynamics

Initial Conditions: Dynamics

Ln(SPILLTECH)t-1

Technological spillovers

0.403 (0.086)

0.282 (0.046)

0.192 (0.037)

0.194 (0.037)

Ln(SPILLSIC)t-1 Product market rivalry

0.044 (0.032)

0.049 (0.031)

0.024 (0.019)

Ln(R&D Stock)t-1 0.495 (0.044)

0.282 (0.046)

0.105 (0.027)

0.104 (0.027)

Ln(Sales)t-1 0.338 (0.052)

0.258 (0.047)

0.138 (0.027)

0.140 (0.027)

Ln(Patents)t-1 0.550 (0.026)

0.550 (0.026)

Pre-sample fixed effect 0.450 (0.049)

0.175 (0.028)

0.174 (0.028)

Over-dispersion (alpha)

0.954 (0.067)

0.814 (0.046)

0.402 (0.029)

0.402 (0.029)

Note: Time dummies and four-digit industry dummies included. Negative binomial model; NT=9,122.

28

Quantification of patent eq

• $1 of R&D associated with 0.007 extra patents (“marginal cost” of patent is about $125,000)

• $1 of SPILLTECH associated with 0.00022 extra patents.

• So $1 of SPILLTECH is worth (in terms of patents) about 3% as much as own R&D

• SPILLSIC does not affect patents

• Higher firm sales associated with more patenting, conditional on R&D (makes patenting more valuable, given innovation)

29

Table 5: R&D Equations (1) (2) (3) (4)

Dependent variable: ln(R&D)

No Effects Fixed Effects Fixed Effects + Dynamics

Fixed Effects + Dynamics

Ln(SPILLTECH) t-1

Technology spillovers 0.224

(0.017) 0.115

(0.071) 0.039

(0.039)

Ln(SPILLSIC) t-1

Product market rivalry

0.291 (0.012)

0.110 (0.026)

0.025 (0.014)

0.030 (0.013)

Ln(Sales) t-1 0.797 (0.009)

0.801 (0.017)

0.218 (0.015)

0.217 (0.015)

Ln(R&D) t-1 0.695 (0.015)

0.695 (0.015)

Ln(Industry Sales) t 0.698 (0.083)

0.133 (0.030)

0.133 (0.022)

0.134 (0.022)

Ln(Industry Sales) t-1 -0.879 (0.083)

-0.085 (0.031)

-0.110 (0.023)

-0.108 (0.022)

Note: Time dummies included; NT=8,565.

30

Quantification of R&D eq

• SPILLSIC and own R&D are positively correlated (implies strategic complementarity). Including sales and firm fixed effects lowers the SPILLSIC effect but it remains significant.

• With fixed effects and dynamics, SPILLTECH does not affect the own R&D decision.

• Both firm and industry sales correlated with own R&D decision

31

Table 6: Production Function (1) (2) (3) Dependent variable: Ln(Sales)

No Fixed Effects

Fixed effects Fixed effects

Ln(SPILLTECH) t-1

Technology spillovers -0.038 (0.009)

0.104 (0.046)

0.111 (0.045)

Ln(SPILLSIC) t-1

Product market rivalry -0.008 (0.004)

0.009 (0.012)

Ln(Capital) t-1 0.291 (0.009)

0.164 (0.012)

0.165 (0.012)

Ln(Labour) t-1 0.646 (0.012)

0.628 (0.015)

0.627 (0.015)

Ln(R&D Stock) t-1 0.059 (0.005)

0.045 (0.007)

0.045 (0.007)

Ln(Industry Sales) t 0.208 (0.040)

0.197 (0.021)

0.198 (0.021)

Ln(Industry Sales) t-1 -0.105 (0.040)

-0.040 (0.022)

-0.040 (0.022)

Note: Time dummies and industry deflators included. NT=10,092

32

Quantification of prod eq

• No fixed effects has negative SPILLSIC (omitted firm specific prices?)

• But with fixed effects only SPILLTEC is significantly and positively related to productivity

• SPILLSIC insignificant

33

Table 7: Theory vs. empirics (under technology spillovers and strategic complementarity)

Partial correlation of: Theory Empirics Consistency?

∂V0/∂rτ Market value with SPILLTECH

Positive .240* Yes

∂V0/∂rm Market value with SPILLSIC

Negative -.067* Yes

∂k0/∂rτ Patents with SPILLTECH

Positive .192* Yes

∂k0/∂rm Patents with SPILLSIC Zero .024 Yes

∂r0/∂rτ R&D with SPILLTECH

Ambiguous .039 -

∂r0/∂rm R&D with SPILLSIC Positive .025* Yes

*significant at 5% level in preferred specifications

34

6. Further investigations

• Industry Heterogeneity – look at 3 high tech sectors (computer hardware, pharma, telecommunications equipment)

• Quantification of spillovers

• Policy Simulations

35

Table 9A. Computer Hardware

A. Computer Hardware (1) (2) (3) (4) (5) Dependent variable Tobin’s Q Patents Cite-

weighted patents

R&D Real Sales

Ln(SPILLTECH)t-1 1.302 (0.622)

0.151 (0.090)

0.338 (0.146)

0.263 (0.199)

0.685 (0.213)

Ln(SPILLSIC)t-1 -0.476 (0.145)

-0.005 (0.153)

0.157 (0.342)

0.039 (0.026)

-0.092 (0.085)

Lagged dependent variable

0.717 (0.065)

0.427 (0.084)

0.684 (0.056)

Observations 358 279 279 390 343

36

Table 9B. Pharmaceuticals

(1) (2) (3) (4) (5) Dependent variable Tobin’s Q Patents Cite-

weighted patents

R&D Real Sales

Ln(SPILLTECH)t-1 1.628 (0.674)

-0.273 (0.326)

1.056 (0.546)

0.407 (0.225)

0.445 (0.208)

Ln(SPILLSIC)t-1 -1.342 (0.612)

-0.106 (0.194)

-0.087 (0.174)

-0.395 (0.452)

-0.391 (0.227)

Lagged dependent variable

0.218 (0.091)

0.269 (0.089)

0.590 (0.147)

Observations 334 265 265 381 313

37

Table 9C. Telecommunications Equipment

(1) (2) (3) (4) (5) Dependent variable Tobin’s Q Patents Cite-

weighted patents

R&D Real Sales

Ln(SPILLTECH)t-1 2.255 (0.870)

0.368 (0.202)

0.658 (0.368)

0.140 (0.246)

0.526 (0.304)

Ln(SPILLSIC)t-1 -0.087 (0.446)

0.036 (0.110)

-0.010 (0.217)

0.033 (0.118)

0.147 (0.156)

Lagged dependent variable

0.590 (0.063)

Observations 405

353 353 429 390

38

Summary on econometric case studies for 3 main high tech sectors

• Evidence of technological spillover effects in all sectors

• Evidence for product market rivalry in computers and pharma, but not in telecom equipment

• Private returns to R&D similar to overall sample ($1.18) in telecom ($1.23), lower in computers ($0.77) and much higher in pharma ($3.65)

• Less clear evidence of strategic complementarity

39

Quantification of spillovers• Calculate long-run equilibrium response of all variables to an

exogenous increase in R&D (Appendix D)• Complex because of multiple linkages between firms through

SPILLTECH and SPILLSIC• Consider first a 1% increase in R&D of all firms and examine

responses in equilibrium values of all variables (Value, Pat, R&D, productivity)

• Distinguish between “autarky” (effects solely from firm changing its own R&D) and “amplification” (include the effects of SPILLTECH and SPILLSIC)

• Main amplification impacts on patents and productivity via SPILLTECH

• From productivity results we see that “social returns” to R&D about 3.5 times higher than private returns

40

Table 8: Impact of a 1% increase in R&D

Variable Amplification Mechanism Autarky Effect Amplification Effect

Total Effect (amplification +

Autarky)

1 R&D 1 0.098(0.053)

1.098(0.053)

2 Patents TECH, SIC and R&D 0.231(0.028)

0.502(0.091)

0.734(0.119)

3 Market Value TECH, SIC and R&D 0.728(0.161)

0.270(0.112)

0.998(0.212)

4 Productivity TECH, SIC and R&D 0.050(0.007)

0.123(0.049)

0.173(0.049)

41

Policy Simulations

• Baseline: 1% R&D shock to all firms• Policy 1: Existing US R&D tax credit• Policy 2: target smaller/medium sized

firms (many EU programs do this)• Policy of targeting small firms yields lower

returns. • Small firms tend to be “less connected”/in

less populated part of productivity space therefore generate less spillover benefits

42

Table 10A: “Policy” simulations (1) (2) (3)

Target Group Total R&D Stimulus, $m

Total R&D Spillovers, $m

Total Productivity Spillovers, $m

1. All Firms 870 95.0 2,717

2. US R&D Tax Credit (firms eligible in median year)

870 94.9 2,747

3. Smaller Firms (smallest 50%) 870 91.2 1,581

4. Larger Firms (largest 50%) 870 95.1 2,767

43

Table 10B, Descriptive statistics. Smaller firms in more niche technology areas

(1) (2) (3)

Target Group % firms SIC TEC

1. All Firms 100 0.046 0.127

2. US R&D Tax Credit (firms eligible in median year)

40 0.052 0.131

3. Smaller Firms (<50%) 50 0.041 0.074

4. Larger Firms (>50%) 50 0.050 0.130

44

7. Conclusions and Extensions• We find evidence of both technological spillovers and

product market rivalry effects of R&D. Results are consistent with predictions of simple analytical framework with strategic complements and technological spillovers.

• Using both technology and product market closeness measures, AND multiple outcome indicators, can help to identify the different effects.

• Useful for analysing impacts of policies – e.g. alternative forms of R&D subsidies

• Extensions – particular sectors; international dimension; specific equilibrium model (Aghion et al, 2005).

45

Back up slides

• Industry heterogeneity (allow all coefficients to vary in PF; case study of computer hardware)

• Deriving the steady state impact on R&D of a shock

46

Our basic 3 equations

47

Re-write in terms of R&D flows

48

First order Taylor series expansion of SPILLTEC term

49

First order Taylor series expansion of SPILLSIC term

50

51

Final “reduced form” of R&D equation

52

Deriving effect of R&D increase on patents

53

From spillover terms in R&D equation From patent equations

Final reduced form for steady state impact on patents