Identifying Technology Spillovers and Product Market Rivalry

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 GrowthCauses and
  Consequences
• San Francisco Federal Reserve Bank
• 18-19 November 2005

2
Introduction

• Two main types of RD spillover effects
• 1. technological spillovers (Griliches, 1992
  Keller,2004)
• 2. product market rivalry effect (I.O. models of
  RD)
• Direct business stealing effect on firm profits
• Strategic effect on RD 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 RD?
• Useful for examining different technology policy
  analysis (e.g. RD 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 RD). Also look at productivity
• Identification of technological spillovers comes
  from the RD, patents and value equations.
• Identification of strategic rivalry effect comes
  from value eq (existence) and RD 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 RD over three times the private return
• See this in pooled sample and in analysis for 3
  high tech sectors
• Simulation of RD policy suggests that medium
  sized firms generate less spillovers than large
  firms (EU policies)

5
Structure

1. Intro
2. Analytical Framework
3. Data
4. Econometrics
5. Results
6. Extensions Industry heterogeneity, policy
   simulations
7. Conclusions

6
2. Analytical Framework

• Two stage game. RD (r) choice at stage 1.
  Knowledge, k (we will proxy by patents) then
  revealed as a function of firms RD. Firms
  choose price/quantity (x) at Stage 2.
• Three firms 0, t and m. Firms 0 and m compete in
  the same product market. Firms 0 and t 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 ?0(k0,km) p
  (k0,x0,xm)
• Firm 0s 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 RD

7
Stage 1 Production of knowledge (k)

• k0 is produced with its own (r0) and firm ts RD
  (rt) k0 F(r0,rt). k0 is increasing in both
  arguments. rt may increase, reduce or leave
  unchanged the marginal product of r0 .
• We analyse how rt and rm affect the best response
  RD, 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 RD
  (rm)
• With any technology spillovers market value rises
  with technological neighbors RD (rt)
• Knowledge, k0 (patents) increases with
  technological neighbors RD rt iff there are
  tech spillovers
• Knowledge is unaffected by rm
• If strategic interaction in product market own
  RD (r0) rises with rival RD (rm) if strategic
  comps and falls with rm if strategic subs

9
Table 1. RD Spillovers and Strategic Rivalry
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/?rt Market value with SPILLTECH Zero Zero Positive Positive
?V0/?rm Market value with SPILLSIC Negative Negative Negative Negative
?k0/?rt Patents with SPILLTECH Zero Zero Positive Positive
?k0/?rm Patents with SPILLSIC Zero Zero Zero Zero
?r0/?rt RD with SPILLTECH Zero Zero Ambiguous Ambiguous
?r0/?rm RD with SPILLSIC Positive Negative Positive Negative
10
Note Ambiguity of effect of tech neighbors RD
on own RD

• ?r0/?rt
• Depends partly on sign impact of neighbours RD
  on marginal product of own RD (?2F/ ? r0? rt)
• Could be positive if own marginal productivity of
  RD increasing in neighbours RD
• Could be negative if fishing out of ideas (own
  marginal productivity of RD decreasing in
  neighbours RD)
• 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 RD, Tobins
  Q, capital, labor and distribution of sales by
  SICs (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 is 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 Tj)/(Ti Ti)1/2(Tj Tj)1/2
  ranges between 0 and 1 for any firm pair i and j
• Technology spillovers are defined as
• SPILLTECHit Sj,j?iTECHi,jGjt
• where Gjt is the RD stock of firm j at time t
  (Gjt Rjt (1-d) Gjt-1)

13
Product Market Spillovers

• Analogous construction of product market
  closeness
• Define Si (Si1, Si2 ,, , Si597)
• where Sik is the of firm is total sales in 4
  digit industry k (k 1,,597)
• SICi,j (Si Sj)/(Si Si)1/2(Sj Sj)1/2
• Product market spillovers from RD are
• SPILLSICit Sj,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
Product market closeness
Technological closeness
16
Examples (High TECH, low SIC) Computer and chip
makers
Table A1
Correlation IBM Apple Motorola Intel
IBM SIC TECH 1 1 0.32 0.64 0.01 0.47 0.01 0.76
Apple SIC TECH 1 1 0.02 0.17 0.01 0.47
Motorola SIC TECH 1 1 0.35 0.46
Intel SIC TECH 1 1
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)
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 RD 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 xs, but also experiment
  with IV/GMM approach)
• Dynamics allow lagged dependent variable
• Demand controls time dummies, industry sales
• Newey-West corrected standard errors

19
(a) Market Value
RD stock/ Fixed assets
Griliches (1981)
Tobins Q (market value to fixed assets)
6th order series expansion (compare with NLLS)
Should be ve
Should be -ve
Fixed effects (long T)
Time dummies
20
(b) RD Equation
RD expenditures
Test of strategic complementarity (ßr2gt0) vs.
strategic substitutes (ßr2lt0)
Compare with Blundell Bond (1998, 2000) GMM-SYS
approach
21
(c) Patent Count Equation
ve
0

• Note Own RD stock in Z D1 if Pgt0, D0,
  otherwise (MFM)
• Allow for overdispersion via Negative Binomial
  model
• Use Blundell, Griffith and Van Reenen (1999)
  control for fixed effects
• with 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

22
(d) Production Function
Deflated sales
ve
0
Own RD, 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

23
Reflection Problem

• Manski (1991) - Identifying endogenous social
  effect (RD 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
  firms distribution of sales across SICs
• Fixed firm effects
• Value function w.r.t. rival RD robust to this
  critique (but RD equation most problematic)
• There remains an issue Do we identify technology
  spillovers or supply shocks?

24
5. Main Results

• Tobins Q equation
• Patents equation
• RD equation
• Productivity equation

25
Table 3 Tobins Q
Dependent 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(RD/Capital) and
time dummies also included. NT10,011
26
Quantification of value eq

• 1 of own RD 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 RD
• 1 of SPILLSIC associated with 4.36 cents lower V
• Industry sales growth raises value, conditional
  on RD variables

27
Table 4 Patent Model
Note Time dummies and four-digit industry
dummies included. Negative binomial model
NT9,122.
28
Quantification of patent eq

• 1 of RD 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 RD
• SPILLSIC does not affect patents
• Higher firm sales associated with more patenting,
  conditional on RD (makes patenting more
  valuable, given innovation)

29
Table 5 RD Equations
Note Time dummies included NT8,565.
30
Quantification of RD eq

• SPILLSIC and own RD 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 RD decision.
• Both firm and industry sales correlated with own
  RD decision

31
Table 6 Production Function
Note Time dummies and industry deflators
included. NT10,092
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/?rt Market value with SPILLTECH Positive .240 Yes
?V0/?rm Market value with SPILLSIC Negative -.067 Yes
?k0/?rt Patents with SPILLTECH Positive .192 Yes
?k0/?rm Patents with SPILLSIC Zero .024 Yes
?r0/?rt RD with SPILLTECH Ambiguous .039 -
?r0/?rm RD 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
36
Table 9B. Pharmaceuticals
37
Table 9C. Telecommunications Equipment
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 RD 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 RD
  (Appendix D)
• Complex because of multiple linkages between
  firms through SPILLTECH and SPILLSIC
• Consider first a 1 increase in RD of all firms
  and examine responses in equilibrium values of
  all variables (Value, Pat, RD, productivity)
• Distinguish between autarky (effects solely
  from firm changing its own RD) 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 RD about 3.5 times higher than
  private returns

40
Table 8 Impact of a 1 increase in RD
Variable Amplification Mechanism Autarky Effect Amplification Effect Total Effect (amplification Autarky)
1 RD 1 0.098 (0.053) 1.098 (0.053)
2 Patents TECH, SIC and RD 0.231 (0.028) 0.502 (0.091) 0.734 (0.119)
3 Market Value TECH, SIC and RD 0.728 (0.161) 0.270 (0.112) 0.998 (0.212)
4 Productivity TECH, SIC and RD 0.050 (0.007) 0.123 (0.049) 0.173 (0.049)
41
Policy Simulations

• Baseline 1 RD shock to all firms
• Policy 1 Existing US RD 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 RD Stimulus, m Total RD Spillovers, m Total Productivity Spillovers, m
1. All Firms 870 95.0 2,717
2. US RD 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 RD Tax Credit (firms eligible in median year) 40 0.052 0.131
3. Smaller Firms (lt50) 50 0.041 0.074
4. Larger Firms (gt50) 50 0.050 0.130
44
7. Conclusions and Extensions

• We find evidence of both technological spillovers
  and product market rivalry effects of RD.
  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 RD 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 RD of a shock

46
Our basic 3 equations
47
Re-write in terms of RD flows
48
First order Taylor series expansion of SPILLTEC
term
49
First order Taylor series expansion of SPILLSIC
term
50
(No Transcript)
51
Final reduced form of RD equation
52
Deriving effect of RD increase on patents
53
Final reduced form for steady state impact on
patents
From spillover terms in RD equation
From patent equations