Public and Private Agricultural R&D Investment and Research Productivity of in China

Transcription

1 Publc and Prvate Agrcultural R&D Investment and Research Productvty of n Chna Yanhong Jn Department of Agrcultural, Food and Resource Economcs Rutgers, The State Unversty of New Jersey 55 Dudley Road New Brunswck, N.J yjn@aesop.rutgers.edu Yahong Hu Department of Agrcultural, Food and Resource Economcs Rutgers, The State Unversty of New Jersey 55 Dudley Road New Brunswck, N.J Carl E.Pray Department of Agrcultural, Food and Resource Economcs Rutgers, The State Unversty of New Jersey 55 Dudley Road New Brunswck, N.J pray@aesop.rutgers.edu Rufa Hu Bejng Insttute of Technology 5 South Zhongguancun Street, Bejng ,Chna rufa@bt.edu.cn Selected Paper prepared for presentaton at the 2016 Agrcultural & Appled Economcs Assocaton Annual Meetng, Boston, Massachusetts, July 31-August 2. Copyrght 2016 by all the authors. All rghts reserved. Readers may make verbatm copes of ths document for non-commercal purposes by any means, provded ths copyrght notce appears on all such copes. 1

2 Publc and Prvate Agrcultural R&D Investment and Research Productvty of n Chna Abstract Employng the count data analyss based on survey data of 1355 frms n Chna s 29 provnces collected n 2007, ths study analyzes the mpact of publc and prvate agrcultural R&D nvestments on research productvty measured by the number of patents granted to agrcultural frms. We fnd that prvate R&D nvestments and havng an own R&D research center ncrease the number of patents granted. However, the publc R&D nvestments do not have a statstcally sgnfcant effects on the number of patents granted. We also fnd that the number of research staff, especally of doctoral research staff, has a postve and statstcally sgnfcant effect on the number of patents granted. Mult-natonal frms and frms located n central Chna have fewer patents than ther counterparts. The man fndngs suggest that t s more effcent for Chnese government to mprove research productvty f t encourages prvate agrcultural R&D nvestments and helps agrcultural frms to buld ther own R&D centers. Chnese government may also need to strengthen the legal framework and nsttutonal resources for the protecton and enforcement of ntellectual propertes to encourage domestc and nternatonal frms patent ther new technologes. Key words: Research and Development Investment, Agrcultural Research Productvty, Publc R&D, Prvate R&D 2

3 Publc and Prvate Agrcultural R&D Investment and Research Productvty n Chna Research and development (R&D) nvestment s consdered as a drvng force of technologcal advances and economc development (Prodan, 2005). In 2000, the global spendng on agrcultural R&D totaled up $36 bllon and about 36% of whch was nvested by the prvate sector. The share of publc and prvate agrcultural R&D nvestment has a strkng dfference between developed and developng countres -- prvate agrcultural R&D nvestment accounted for 93% n developed countres, but only 6% n developng countres (Pardey et al. 2006). The lack of prvate agrcultural R&D nvestments n developng countres s manly due to weak ntellectual property rghts, government control of agrcultural nput markets, and lmted foregn drect nvestment (Pray and Fugle, 2002). However, the prvate sector plays an ncreasngly mportant role, especally n food processng research and development (Fugle, et al., 2011, Fugle and Toole, 2014). Publc and prvate agrcultural R&D can potentally affect research productvty dfferently due to ther dstnct focuses. Publc R&D nvestments manly focus on basc and appled research, whle prvate R&D largely focuses on experment research. Prevous research has shown the postve effect of publc agrcultural research on total factor productvty (TFP) growth n agrculture (Alston, et al., 2009, Huffman and Evenson, 2008, Wang, et al., 2012). The lterature s scarce n addressng the role of R&D nvestments on research productvty, let alone dstngushng publc and prvate R&D nvestments. Among varous measures of research productvty, the number of patents granted s commonly used to measure technologcal and scentfc nnovatons due to the fact that patent data are more readly accessble (Grlches, 1984). The objectves of ths study s to examne the productvty of publc and prvate agrcultural R&D nvestments n Chna. Employng the count data analyss based on survey data of 1355 frms n Chna s 29 provnces collected n 2007, ths study analyzes the mpact of publc and prvate agrcultural R&D nvestments on research productvty measured by the number of patents granted to agrcultural frms. We fnd 3

4 that prvate R&D nvestments and havng an own R&D research center ncrease the number of patents granted. However, the publc R&D nvestments do not have a statstcally sgnfcant effects on the number of patents granted. We also fnd that the number of research staff, especally of doctoral research staff, has a postve and statstcally sgnfcant effect on the number of patents granted. Mult-natonal frms and frms located n central Chna have fewer patents than ther counterparts. The man fndngs suggest that t s more effcent for Chnese government to mprove research productvty f t encourages prvate agrcultural R&D nvestments and helps agrcultural frms to buld ther own R&D center. Strengthen the legal framework and nsttutonal resources for the protecton and enforcement of ntellectual propertes also helps to encourage domestc and nternatonal frms patent ther new technologes. Lterature Revew The mpact of R&D nvestment on productvty has been wdely addressed n the lterature (Alene, 2010, Balcombe, et al., 2005, Block, 2010, Salm and Islam, 2010). Po-Ch, et al. (2008) frst examne the productvty growth of Chna n They fnd that techncal progress postvely contrbutes to the ncreasng growth rate of productvty, and the publc R&D nvestment s one of the mportant ncentves for technologcal progress. Instead of agrcultural productvty, ths study focuses on research productvty resultng from R&D nvestments. Research productvty can be measured by ether the number of patents (Grlches, 1994, Prodan, 2005) or the number of patents per unt of the R&D nvestment (Lanjouw and Schankerman, 2002). De Rassenfosse and de la Pottere (2009) clam that research efforts lead to nventons and nventons lead to patents. Inventons are most trggered by productvty effects, whereas patents are caused by the propensty to patent effect. Thus, patent counts could be consdered ether as an ndcator of propensty to patents or an ndcator of research productvty. Han and Lee (2007) fnd a postve assocaton between R&D nvestments and the number of patent per employee. Jaffe (1989) fnd a postve effect of unversty research on 4

5 the number of patents granted to frms. Branstetter and Sakakbara (2000) collect data on all companyto-company cooperatve R&D projects formed wth a degree of government nvolvement from 1982 to They fnd that the Japanese government sponsored R&D ncreases the number of patents that a frm owned. It s common to have a tme delay between the tme a patent applcaton s submtted and s granted. For example, t takes 2-9 years for a patent to be granted after ts applcaton s submtted (Kondo, 1999). Some studes use the number of patent applcatons to measure research productvty of R&D nvestments (De Rassenfosse and de la Pottere, 2009). Yet, the lterature s scarce to separate publc and prvate R&D nvestments when examnng the effect of each on research productvty. Ths study wll fll the gap by dstngushng publc and prvate R&D nvestment and further separatng the publc R&D nvestment by those focusng on appled research and those on development. Methodology We employ count data modelng approaches to nvestgate the effects of both publc and prvate agrcultural R&D nvestments on the number of patents granted to agrcultural frms. Appendx A presents the four most popular count data models, namely, Posson, Negatve Bnomal (NB), zeronflated Posson (ZIP), and zero-nflated NB (ZINB) (Cameron and Trved, 2013). Posson model assumes that varance equals to ts mean, whch can lead to neffcent estmates f over-dsperson s present n the count data. Over-dsperson s caused by ether unobserved heterogenety among ndvduals or excess zeros n the dependent varable. When unobserved heterogenety s a concern, a NB model has been suggested and t adds an error term to the condtonal mean of the Posson dstrbuton. Both Posson and NB models do not account for excess zeros and thus can produce based estmates. Excess zeros can be a concern n ths study because more than 68% of the observatons has no patent as all (see Table 1). Zero-nflated regresson models, such as ZIP and ZINB models, are warranted to address the ssue of havng excess zeros. Both ZIP and ZINB nclude 5

6 a logt (or probt) regresson for zero nflaton, followed by the Posson estmaton for ZIP or the negatve bnomal estmaton for ZINB. Based on statstcal tests on the null hypothess 0 for over-dsperson for nested models (Posson vs. NB, and ZINB vs. ZIP) and the Vuong test for nonnested models (ZINB vs NB, and ZIP vs. Posson), we are able to choose the most sutable one among these four count data models. As shown n Fgure 1, f the Vuong test favors the ZINB model over the NB model, then a statstcal test on 0 s conducted to contrast ZINB versus ZIP. If 0 s rejected, ZINB s the most approprate specfcaton, and both ndvdual heterogenety and excess zeros contrbute to the overdsperson. Otherwse, ZIP model s compared to Posson model by usng the Vuong test. If ZIP s the most approprate specfcaton, then only excessve zeros account for over-dsperson. Otherwse no over dsperson s present and Posson s favored. On the other hand, f the Vuong test favors the NB model, then we wll test f the heterogenety parameter s sgnfcantly dfferent from zero to contrast NB vs. Posson. A rejecton of 0 suggests that the NB model s most approprate specfcaton and heterogenety accounts for over-dsperson. Otherwse, the Posson and ZIP are compared. Data and Varable Constructon The data set used for ths study come from a naton-wde mal survey of agrbusness frms n 29 provnces (Hebe and Tbet are not ncluded) n Chna. The survey was ntated by the Mnstry of Agrculture and mplemented by CCAP (Center for Chnese Agrcultural Polcy) n It collected nformaton of agrcultural R&D nvestments, government subsdes for agrcultural research, frm attrbutes, and R&D research centers/dvsons n 2000, 2004, 2005 and The respondent reported the total number of patents granted by There were 503, 1059, 1236, and 1365 frms n the year of 2000, 2004, 2005 and 2006, respectvely. We excluded 10 observatons n the machnery, pestcde, or fertlzer ndustres that were overseen by the Mnstry of Agrculture. The remanng 6

7 1355 frms are classfed nto four ndustres: crop, lvestock, food processng, and fshery. As shown n Table 1, more than 2/3 dd not have any patent (N = 927) and the majorty of frms have ether one or two patents. More than half were process patents, followed by new product patents (38.02%), and the least for packagng and marketng patents (11.71%) such dstrbuton pattern stll hold f we examne the number of patents by ndustry (Fgure 1). As shown n Fgure 1, the greatest share of process patents s found n the fshery ndustry (79.37%), product patents for the lvestock ndustry (48.73), and packagng and marketng patents for the food process ndustry. Table 2 provdes the summary statstcs of the key varables. Except the number of patents, the other varables take the annual average n 2000, 2004, 2005 and All the monetary values are deflated by consumer prce ndex of Prvate R&D nvestments consst of those of own nvestments, through contracts, or receved from other frms. Prvate agrcultural R&D nvestment was more than doubled -- ncreasng from 0.74 mllon Yuan n 2000 to 1.61 mllon Yuan n About half of frms have an n-house R&D center, less than 1% have R&D nvestments through contracts, and approxmately 13-14% have both an n-house R&D nvestment center and through contracts. The government subsdy ncreased from mllon Yuan n 2000 to mllon Yuan n We dsaggregate the publc R&D nvestment nto those nvested n appled research and expermental development (Publc-R) and those nvested n basc research (Pubc-D). Accordng to the Frascat Manual (2002), Basc research s expermental or theoretcal work undertaken prmarly to acqure new knowledge of the underlyng foundatons of phenomena and observable facts, wthout any partcular applcaton or use n vew (Frascat Manual 2002, p.77). Appled research s orgnal nvestgaton undertaken n order to acqure new knowledge. It s, however, drected prmarly towards a specfc practcal am or objectve (Frascat Manual 2002, p.77). Expermental development s systematc work, drawng on knowledge ganed from research and practcal experence, whch s drected to producng new materals, products and devces; to nstallng new 7

8 processes, systems and servces; or to mprovng substantally those already produced or nstalled (Frascat Manual 2002, p. 77). Thus, the focus of Publc-R and Publc-D nvestments dffer. They are expected to affect research productvty dfferently. Despte the dramatc ncrease of prvate agrcultural R&D nvestment, t was stll outweghed by publc R&D nvestment. We also compare some mportant varables between frms wth and wthout patents. As shown n Table 3, compared wth frms wth no patents, frms wth patents granted have hgher sales revenue and research staff and they also have greater government subsdes for research and publc and prvate agrcultural R&D nvestment. Estmaton Results We assume that the number of patents granted s affected by both prvate and publc R&D nvestments, government subsdes for research, the qualty of research staff, and frms attrbutes such as ownershp, age, and sales revenue. Publc R&D nvestment s dvded nto the nvestment focusng on appled research and development separately. The R&D nvestment varables and the government subsdes for research take the average value n three years at the frm level. Whether a frm has an n-house R&D center s also ncorporated. The human captal of R&D actvtes measured by total number of research staff wthout a PhD degree and total number of research staff that hold a PhD degree are ncorporated separately n the model. We also control for the dfference by regon and ndustry. Two man hypotheses wll be emprcally tested. Hypothess 1: Prvate R&D nvestment and havng an own n-house R&D center ncreases the patent count. Hypothess 2: Publc R&D nvestment on appled research and government subsdes for research has a postve mpact whle publc R&D nvestment on development has a negatve mpact on research output measured by patent counts. 8

9 STATA s used to estmate the four models. The estmaton results are reported n Appendx B. Table 4 present the margnal effects of the four models and the dscussons are based on the margnal effects. As shown n Table 4, the Vuong test (Vuong-statstc = 5.04 and p-value = 0.00) suggests that the ZINB model fts the data better than the NB model, and the lkelhood-rato test of 0 ndcates that the ZIP model outperforms the ZINB model (LR-statstc = 0.24). Furthermore, the ZIP model fts the data better than the Posson model based on the lkelhood rato test (LR-statstc = 9.94 and p-value = 0.00). We therefore conclude that the ZIP model s a more approprate specfcaton than Posson, NB, or ZINB models. Ths concluson s also renforced by the fndng that, relatve to the other pooled models, ZIP has the hghest rato of correct predctons (43.08%). The man results are summarzed below. Frst, prvate R&D nvestments and havng ts own R&D research center have a statstcally sgnfcant, postve mpact on the number of patents granted. Ths fndng support Hypothess 1 and t s also consstent wth the prevous research (Grlches, 1984; Jaffe, 1989; Kondo, 1999; Han and Lee, 2007; De Rassenfosse, 2009). The effect of the n-house R&D research center can be explaned below. Frms that are nclned to patent the new technologes prefer n-house R&D, whch provdes better protecton of the ntellectual propertes before the patents are granted. Second, as we expected, publc R&D nvestment on appled research has a postve mpact, but publc R&D nvestment on development has a negatve effect. But these effects are not statstcally sgnfcant. On the other hand, government subsdy for prvate research has a postve mpact on the number of patents granted but s not statstcally sgnfcant. The fndngs partally support Hypothess 2. Thrd, foregn frms are granted sgnfcantly fewer patents than state-owned frms n Chna. We provde two explanatons for ths fndngs. Frst, most of foregn frms cooperated wth prvate frms or state-owned frms n Chna, and thus ths type of frms are classfed nto jont ownershp frms. Among the 1355 frms, only 2% were foregn frms. Another reason can be that 9

10 foregn-owned frms prefer applyng patents outsde of Chna where more adequate ntellectual property laws are enforced. We also fnd that human captal measured by the number of research staff has a postve, statstcally sgnfcant assocaton wth the number of patents. The result s consstent wth the prevous lterature (Han and Lee, 2007). Furthermore, we ncorporate the number of research staff who have a Ph.D. degree and those havng no Ph.D. degree n the models. We fnd that the effect of the Ph.D. research staff s greater than that of the research staff wthout a Ph.D. degree. Frms located n the east and west regons have more patents than frms n the central regon, and frms n the east Chna have the most patent counts than frms n other two regons. Conclusons Publc sectors had domnated the agrculture research n Chna, and prvate sectors started makng an mportant role on technology nnovaton and productvty growth untl recently after the polcy reforms. Usng survey data on 1355 frms across 29 provnces across the year of 2000, 2004, 2005 and 2006 n Chna, ths study analyzes the mpact of publc and prvate R&D nvestment on research productvty measured by the number of patents. We fnd a strong and postve relatonshp between prvate R&D nvestments and the number of patents granted. Frms wth both own R&D centers and obtan technology through contracts owned the most patents, followed by frms wth only own R&D center, and least for those who have no R&D actvtes. However, the publc R&D nvestments do not have a statstcally sgnfcant effects on the number of patents granted. We also fnd that the number of research staff, especally of doctoral research staff, has a postvely sgnfcant effect on the number of patents granted. Thrd, mult-natonal frms and frms located n central Chna have fewer patents than ther counterparts. Ths study offers several polcy mplcatons. It s more effcent for Chnese government to mprove research productvty f the government encourages prvate agrcultural R&D and helps 10

11 agrcultural frms buld ther own R&D centers. Strengthen the legal framework for the protecton and enforcement of ntellectual propertes s also mportant to encourage domestc and especally nternatonal frms patent ther new technologes. Reference Alene, A.D "Productvty growth and the effects of R&D n Afrcan agrculture." Agrcultural Economcs 41: Alston, J.M., et al Persstence pays: US agrcultural productvty growth and the benefts from publc R&D spendng: Sprnger Scence & Busness Meda. Balcombe, K., A. Baley, and I. Fraser "Measurng the mpact of R&D on productvty from a econometrc tme seres perspectve." Journal of Productvty Analyss 24: Block, S. "The declne and rse of agrcultural productvty n sub-saharan Afrca snce 1961." Natonal Bureau of Economc Research. Branstetter, L.G., and M. Sakakbara. "When do research consorta work well and why? Evdence from Japanese panel data." Natonal Bureau of Economc Research. Cameron, A.C., and P.K. Trved Regresson analyss of count data: Cambrdge unversty press. De Rassenfosse, G., and B.v.P. de la Pottere "A polcy nsght nto the R&D patent relatonshp." Research Polcy 38: Fugle, K., et al "Research nvestments and market structure n the food processng, agrcultural nput, and bofuel ndustres worldwde." USDA-ERS Economc Research Report. Fugle, K.O., and A.A. Toole "The Evolvng Insttutonal Structure of Publc and Prvate Agrcultural Research." Amercan Journal of Agrcultural Economcs:aat107. Greene, W.H "Accountng for excess zeros and sample selecton n Posson and negatve bnomal regresson models." 11

12 Grlches, Z "Productvty, R&D, and the data constrant." The Amercan Economc Revew 84:1-23. Han, Y.-J., and W.-Y. Lee "The effects of the characterstcs of Korean frms on the patent producton functon." Economcs of Innovaton and New Technology 16: Huffman, W.E., and R.E. Evenson Scence for agrculture: A long-term perspectve: John Wley & Sons. Kondo, M "R&D dynamcs of creatng patents n the Japanese ndustry." Research Polcy 28: Lambert, D "Zero-nflated Posson regresson, wth an applcaton to defects n manufacturng." Technometrcs 34:1-14. Lanjouw, J.O., and M.A. Schankerman "Research productvty and patent qualty: measurement wth multple ndcators." Po-Ch, C., et al "Total factor productvty growth n Chna's agrcultural sector." Chna Economc Revew 19: Pray, C., and K. Fugle "Prvate Investment n Agrcultural Research and Internatonal Technology Transfer n Asa, Bogor, Indonesa." Agrcultural Economc Report 805:155. Prodan, I "Influence of Research and Development expendtures on number of patent applcatons: selected case studes n OECD countres and central Europe " Appled Econometrcs and Internatonal Development 5:5-22. Salm, R.A., and N. Islam "Explorng the mpact of R&D and clmate change on agrcultural productvty growth: the case of Western Australa*." Australan Journal of Agrcultural and Resource Economcs 54: Vuong, Q.H "Lkelhood rato tests for model selecton and non-nested hypotheses." Econometrca: Journal of the Econometrc Socety:

13 Wang, S., et al "Accountng for the Impacts of Publc Research, R&D Spll-ns, Extenson, and Roads n US Agrcultural Productvty Growth." Agrcultural Productvty: An Internatonal Perspectve. Walngford, UK: CABI. 13

14 Table 1. Dstrbuton of the number of patents granted to agrcultural frms No. of patents No. of frms % of frms By patent type No. of frms No. of patents % of patents Product Process Packagng & Marketng Total Total Source: Calculated by the author based on the CCAP survey

15 Table 2 Summary Statstcs of Key Varables Varable No. of observatons Total Prvate R&D nvestment (Mllon yuan) 0.74 (2.71) 1.06 (3.61) 1.25 (3.89) 1.61 (4.59) Patent number (count data) 0.96 (1.66) 0.88 (1.58) 0.82 (1.53) 0.80 (1.52) Publc-R (Mllon yuan) (67.44) (77.49) (92.36) (103.96) Publc-D (Mllon yuan) (170.41) (340.48) (386.70) (430.62) Government subsdy (Mllon yuan) (0.358) (0.722) (0.780) (0.601) Sale revenues (Mllon yuan) (238.64) (427.28) (499.59) (594.85) Frm age (years) 7.53 (8.86) 7.20 (7.37) 7.33 (7.16) 7.79 (7.08) PhD R&D Staff (%) 0.13 (0.69) 0.18 (0.84) 0.24 (1.04) 0.34 (1.31) Publc lsted company (yes = 1) 0.02 (0.15) 0.02 (0.14) 0.02 (0.13) 0.02 (0.13) Ownershp: Prvate 0.61 (0.49) 0.70 (0.46) 0.72 (0.45) 0.73 (0.44) State 0.16 (0.37) 0.11 (0.31) 0.09 (0.29) 0.09 (0.29) Foregn (0.04) (0.03) (0.049) (0.05) Other 0.09 (0.51) 0.06 (0.43) 0.06 (0.42) 0.06 (0.42) Collectvely-owned 0.13 (0.33) 0.12 (0.32) 0.11 (0.31) 0.11 (0.31) Sector: Crops 0.25 (0.44) 0.26 (0.44) 0.27 (0.45) 0.27 (0.44) Lvestock 0.25 (0.44) 0.24 (0.43) 0.24 (0.43) 0.25 (0.43) Fshery (0.25) Food Processng 0.42 (0.50) R&D dvson (dummes) R&D dvson: In House R&D 0.54 (0.50) Contract R&D 0.06 (0.24) In-house & contract R&D 0.14 (0.34) No R&D 0.26 (0.44) (0.25) 0.43 (0.50) 0.52 (0.50) 0.08 (0.26) 0.14 (0.34) 0.27 (0.44) (0.25) 0.43 (0.50) 0.50 (0.50) 0.07 (0.27) 0.14 (0.34) 0.28 (0.45) (0.24) 0.42 (0.49) 0.50 (0.50) 0.08 (0.27) 0.13 (0.34) 0.29 (0.45) 15

16 Table 3 Comparson of key varables between frms wth and wthout patents Frms wth no patent Frms wth at least one patent Number of Observatons Average patent count Sales revenue (1,000,000 Yuan) Prvate R&D nvestment (1,000,000 Yuan) Research Staff wthout a PhD degree Research Staff wth PhD degree Government Subsdes for prvate research (1,000,000 Yuan) Publc R (1,000,000 Yuan) Publc D (1,000,000 Yuan) Source: Calculated by the author based on the CCAP survey

17 Table 4. Margnal Effects based on the Four Count Data Models Varable Posson NB ZIP ZINB Prvate R&D nvestment 0.024*** 0.027*** 0.036*** (Mllon yuan) (0.004) (0.007) (0.014) (0.043) Publc-R (Mllon yuan) (0.013) (0.010) (0.014) (0.013) Publc-D (Mllon yuan) (0.003) (0.002) (0.004) (0.003) Government subsdy (Mllon yuan) 0.032*** 0.019*** (0.003) (0.015) (0.057) (0.099) Sale revenues (Mllon yuan) (0.0007) (0.0001) (0.0006) (0.0001) Frm age (years) (0.004) (0.005) (0.006) (0.006) Non PhD R&D Staff *** ** 0.004** (0.0001) (0.001) (0.001) (0.002) PhD R&D Staff 0.013*** 0.023*** 0.031*** 0.036* (0.003) (0.007) (0.008) (0.019) Publcally traded frms ** ** (0.151) (0.175) (0.189) (0.207) Ownershp (base=state-owned) Prvate (0.181) (0.099) (0.240) (0.236) Foregn partcpaton ** * (0.233) (0.393) (0.191) (0.234) Collectvely-owned (0.104) (0.092) (0.159) (0.147) Other (0.117) (0.173) (0.133) (0.143) Sector (base=crops) Lvestock (0.394) (0.218) (0.446) (0.418) Fshery (0.132) (0.137) (0.140) (0.110) Food Processng (0.156) (0.112) (0.152) (0.163) R&D dvson (base=own R&D) Contract R&D *** *** *** *** (0.058) (0.068) (0.074) (0.061) Both R&D ** 0.225* (0.107) (0.087) (0.133) (0.160) No R&D *** *** *** *** (0.006) (0.009) (0.006) (0.027) Regon (base=central) East 0.156*** 0.154*** 0.182*** 0.193*** (0.012) (0.023) (0.017) (0.009) West 0.073*** 0.089*** 0.122*** 0.133*** (0.009) (0.022) (0.019) (0.004) Overall predcton accuracy 41.71% 42.08% 43.08% 42.30% Vuong test ZIP vs. Posson: Z=9.94 ZINB vs. NB: Z=5.04 Αlpha test NB vs. Posson: α=1.97 ZINB vs. ZIP: α=0.24 Note: The astersk, *, **, and *** ndcates 10%, 5% and 1% sgnfcance level, respectvely. 17

18 ZINB vs. NB Vuong test favor ZINB favor NB ZINB vs. ZIP Posson vs. NB test on ZIP vs. Posson test on fal to reject fal to reject Vuong test reject reject favor ZIP favor Posson Fgure 1: Procedure to choose an approprate model among the Posson, NB, ZIP and ZINB models 18

19 crops lvestock fshery food processng product process package Fgure 2. Dstrbuton of patents by ndustry sector and patent type 19

20 Appendx A. A Bref Revew of Four Count Data Models In a basc Posson regresson model wth a logarthm lnk functon, the number of events y for ndvdual has a Posson dstrbuton wth a condtonal mean dependng on ndvdual s characterstcs, x : (1) E y exp x x x, where s a vector of unknown coeffcents assocated wth the covarate vector x. For convenence of notaton, we drop x n x and use below. The probablty densty functon of y gven x s (2) f y x exp y. y! The NB model adds an error term, to the condtonal mean of the Posson dstrbuton to model the unobserved heterogenety, (3) E y exp x x. where exp s normally assumed to follow a gamma dstrbuton wth mean one and varance. The probablty densty functon of y gven x now becomes 1/ ( y 1/ ) (4) 1 f y x y! (1/ ) 1 1 / The condtonal mean and varance of y under the NB model are E and y x (5-1) VAR y (1 ) (5-2) x. where s the varance of gamma dstrbuton and ndcates the degree of over-dsperson. As becomes larger, the dstrbuton wll be more dspersed. As gets close to zero, the NB model converges to the Posson model. The Posson and NB models are nested, and a statstcal rejecton of the null hypothess of 0 wll favor NB over Posson specfcaton. (6) Lambert (1992) frst ntroduced ZIP model as y 0 y ~ Posson wth probablty π - π wth probablty 1 ( y 0,1,2,...) The probablty of havng an extra zero whch s not subject to the Posson dstrbuton, π, s assumed to have a logt functon (7). The unobserved probablty s generated as a logstc or probt functon y. 20

21 of observable covarates to ensure nonnegatvty. The choce between logt and probt s usually unmportant snce the two functons are smlar and usually gve very smlar results. (7) z exp, 1 exp z where z s a vector of observable covarates and s a vector of coeffcents assocated wth z. The mean and varance of y n the ZIP model are E (1 ) and y x (8-1) (8-2) VAR 1 (1 ). y x Equatons (8-1) and (8-2) show that ndcates the degree of over-dsperson. As π approaches 1 zero, the ZIP model converges nto the Posson model. Smlarly, to account for ndvdual heterogenety and excess zeros smultaneously, ZINB model wth a logt lnk functon s used. The mean and varance of y under the ZINB model are E (1 ) and y x (9-1) (9-2) VAR 1 1. y x Equatons (9-1) and (9-2) show that reflects the degree of over-dsperson n the ZINB models, 1 whch accounts for over-dsperson from both zero nflaton and unobservable heterogenety. The Posson and ZIP models are not nested, and nether are the NB and ZINB models. Vuong (1989) proposed a lkelhood rato test for non-nested models, and Greene (1994) adapted the technque for the cases of ZIP versus Posson, and ZINB versus NB models. The test statstc s (10) Z N m, s m where m and s m are the mean and standard devaton of m and N s the number of observatons. m p y x s defned as m ln 1 where p y x p y x 1 and p 2 y x are the predcted probabltes from 2 the competng models. Asymptotcally, Z has a standard normal dstrbuton, wth large postve values (>1.96) favorng the zero-nflated model and wth large negatve values (<-1.96) favorng the nonzeronflated model at a 5% sgnfcance level. 21

22 22

23 Appendx B: Estmaton Results of the Four Count Data Models Zp ZINB Varable Posson NB ZIP ZIP Inflated ZINB Inflated Prvate R&D nvestment 0.044*** 0.049*** 0.020** (Mllon yuan) (0.008) (0.013) (0.013) (0.053) (0.024) (0.160) Publc-R (Mllon yuan) (0.023) (0.019) (0.035) (0.037) (0.045) (0.055) Publc-D (Mllon yuan) (0.006) (0.004) (0.008) (0.009) (0.011) (0.013) Government subsdy 0.058*** ** (Mllon yuan) (0.005) (0.009) (0.038) (0.205) (0.069) (0.388) Sale revenues (Mllon yuan) (0.066) (0.148) (0.067) (0.0002) (0.104) (0.0005) Frm age (years) (0.008) (0.010) (0.004) (0.010) (0.004) (0.012) Non PhD R&D Staff *** 0.001*** *** ** (0.000) (0.001) (0.000) (0.003) (0.001) (0.004) PhD R&D Staff 0.024*** 0.043*** (0.006) (0.013) (0.051) (0.089) (0.057) (0.145) Publcally traded frms (0.467) (0.630) (0.284) (1.320) (0.314) (1.493) Ownershp (base=state owned) Prvate ** ** * (0.296) (0.168) (0.150) (0.315) (0.206) (0.243) Foregn partcpaton *** ** (1.325) (1.546) (0.763) (4.941) (0.954) (2.139) Collectvely-owned ** ** (0.187) (0.167) (0.072) (0.316) (0.113) (0.297) Other (0.215) (0.267) (0.232) (0.169) (0.268) (0.231) Sector (base=crops) Lvestock (0.618) (0.359) (0.915) (0.988) (1.007) (1.121) Fshery (0.254) (0.266) (0.321) (0.744) (0.300) (0.946) Food Processng (0.273) (0.196) (0.257) (0.603) (0.256) (0.809) R&D dvson (base=own R&D) Contract R&D *** *** *** *** (0.301) (0.3331) (0.213) (0.720) (0.231) (0.836) Both R&D ** *** * (0.160) (0.126) (0.108) (0.199) (0.142) (0.462) No R&D *** *** *** 1.558*** ** *** (0.051) (0.049) (0.240) (0.291) (0.35) (0.243) Regon (base=central) East 0.280*** 0.278*** 0.151*** *** *** (0.024) (0.038) (0.027) (0.050) (0.027) (0.055) West 0.129*** 0.158*** *** *** *** (0.017) (0.040) (0.027) (0.028) (0.029) (0.156) Note: The astersk, *, **, and *** ndcates 10%, 5% and 1% sgnfcance level, respectvely. 23