¹ Especially after 1995, policies were designed to accelerate indigenous science and technology development. The number of patent applications from domestic applicants was 10 011 in 1995, and it has increased dramatically ever since. The number reached 39 806 by 2002 and 255 832 by 2011 (State Intellectual Property Office of People's Republic of China, 2012). See http://www.english.sipo.gov.cn/.

² Indigenous innovation means “to develop the capability to conduct R&D or create innovation internally”.

³ For more information on the ICS, see http://www.worldbank.org.

⁴ For more detailed information, the reader is referred to http://en.wikipedia.org/wiki/Economy_of_the_People%27s_Republic_of_China.

⁵ Source: Ministry of Science and Technology of the People's Republic of China (http://www.most.gov.cn/eng/).

⁶ Main source: OECD library (http://www.oecd-ilibrary.org/economics/oecd-factbook-2005_factbook-2005-en).

⁷ Index b refers to “buy”, i.e. technology purchases; index m refers to “make”, i.e. own R&D.

⁸ Among these six cases, two are in the electronic equipment industry, two in the electronic-parts-making industry, one in the garment and leather products industry and one in the metallurgical products industry.

⁹ We prefer to use the dynamic panel GMM approach rather than the proxy-based approach introduced by Olley & Pakes (1996 Olley, S. and Pakes, A. 1996. The dynamics of productivity in the telecommunications equipment industry. Econometrica, 64: 1263–1298. [Crossref], [Web of Science ®] , [Google Scholar]). Ackerberg et al. (2006 Ackerberg, D., Caves, K. & Frazer, G. (2006) Structural identification of production functions, mimeo, R&R Econometrica  [Google Scholar]) compare the two approaches and discuss their respective advantages and disadvantages.

¹⁰ See footnote 3.

¹¹ According to the National Bureau of Statistics of China, small-sized firms have fewer than 300 employees, medium-sized firms have between 300 and 2000 employees and large-sized firms have more than 2000 employees. The classification is based on the number of long-term employees according to “The classification of small, medium and large Chinese manufacturing firms” from the National Bureau of Statistics of China (http://www.stats.gov.cn/). The large-sized firms included in the sample, after dropping missing values, represent about 3.5% of the total sample (33 firms). We decided to drop these firms and look only at the small- and medium-sized firms. Meanwhile, there are about 40% of firms with fewer than 100 employees. It is interesting to look at their technology-sourcing behaviour separately, since the t-tests in Tables 2 and 4 show that their characteristics are significantly differently from those of other firms. In total, our sample is divided into three categories: firms with fewer than 100 employees (40.1%), firms with a number of employees between 100 and 300 (30.37%), and firms with more than 300 employees (29.53%).

¹² Firms with no information on financial outcomes are omitted. The service sector is not included because the innovation outputs are not reported. Furthermore, following Hall & Mairesse (1995 Hall, B. H. and Mairesse, J. 1995. Exploring the relationship between R&D and productivity in French manufacturing firms. Journal of Econometrics, 65(1): 263–293. [Crossref], [Web of Science ®] , [Google Scholar]), we keep observations only for cases in which the capital–labour ratio is within three times the inter-quartile range (the difference between the 75% value and the 25% value) above or below the median. This removed 129 observations, or 2.1% of the sample. The remaining sample is an unbalanced panel with 3332 observations for the period 2000–2002.

¹³ For the annual sales and profits, we use the wholesale-price deflators at the industry level. For capital, material and innovation expenditure, the industry-input deflators are used (National Bureau of Statistics of China, 2000–2003).

¹⁴ The capital and material inputs are logarithmically transformed and expressed in 1000 RMB per employee.

¹⁵ The percentage of firms that claimed to have positive expenditures of R&D and TP in the period under review.

¹⁶ The Delta method is based on the first-order approximation around the mean. For more details, see Oehlert (1992 Oehlert, G. W. 1992. A note on the delta method. American Statistician, 46: 27–29. [Taylor & Francis Online], [Web of Science ®] , [Google Scholar]).

¹⁷ When computing the marginal effects of I _b or I _m , we take into account all the terms where these two dummies appear (i.e. four terms in total).

¹⁸ When computing the marginal effects of I _m  ln(R&D/labour) and I _b  ln(TP/labour), the elasticities of the square term (I _m  ln(R&D/labour)²/2 and I _b  ln(TP/labour)²/2) and the interaction term (I _m  ln(R&D/labour) × I _b  ln(TP/labour)) are also accounted.

¹⁹ As Ai & Norton (2003 Ai, C. and Norton, E. 2003. Interaction terms in logit and probit models. Economics Letters, 80: 123–129. [Crossref], [Web of Science ®] , [Google Scholar]) pointed out, complementarity should be based on a positive sign of the cross-elasticity and not the interaction term alone.

²⁰ The number of orthogonality conditions equals the number of instruments used for the endogenous variables plus the number of pre-determined variables (used as their own instruments). In the GMM case, 14 variables are considered to be endogenous (the first 14 independent variables in Table 6). We consider the first differences in the error term of the productivity equation to be orthogonal to the one- and two-period lagged levels of the endogenous variables for the year 2002 and to the one-period lagged levels of the endogenous variables for the year 2001 (remember, we have no data prior to 2000). That makes 42 orthogonality conditions so far. As to the level of the error term, we consider it to be orthogonal to the first differences in the endogenous variables for 2002 and 2001. That makes 28 additional orthogonality conditions (14 for each year). So, in total, the system GMM exploits 70 orthogonality conditions to instrument for the endogenous variables, plus 15 orthogonality conditions using the predetermined variables of our model (the time, industry, ownership and exporting status dummies) in the first difference equations, which makes a grand total of 85 orthogonality conditions. The number 56 in the Hansen overidentification test equals the total number of orthogonality conditions (85) minus the number of estimated parameters (29).

²¹ The two-step system GMM performed in STATA 11, using the xtabond2 command, automatically corrects the standard errors for heteroskedasticity and autocorrelation (Roodman, 2006 Roodman, D. (2006) How to do xtabond2: an introduction to “difference” and “system” GMM in Stata, Working Paper Number 103, Centre for Global Development  [Google Scholar]).

²² By dividing all the inputs by labour and dividing output by labour, we have imposed that the second-order parameters of the translog approximation sum up to 0 and the first-order terms sum up to 1, so as to impose constant returns to scale. We have then allowed the sum of the first-order terms to deviate from 1 by allowing for a labour first-order term to appear on the right-hand side.

²³ When the marginal effects were computed, we included the elasticities of the square term and the interaction term. See footnotes 17 and 18. For computing the marginal effects of ln(capital/labour) and ln(material/labour), the square term and the interaction term are also included.

²⁴ We thank an anonymous referee for this remark.