I. Introduction

Does the stock market affect a firm’s innovation activity? This is an important question from both macro and micro perspectives. At the macro level, technological innovation is vital for a country’s economic growth (Schumpeter, 1943; Solow, 1957; Aghion and Howitt, 1992). At the micro level, innovation capacity determines a firm’s long-term competitive advantage (Porter, 1992).

On the one hand, a stock market is important for firms to raise capital. An initial public offering (IPO) can help increase a firm’s innovation activity by providing access to a funding source of lower cost than debt financing (Hall and Lerner, 2010). Moreover, as Holmstrong (1989) points out, the payoff of innovation is heavily skewed and risky due to its long-term, idiosyncratic nature. As a result, debt is a less efficient way of financing innovation compared to equity. For these reasons, an IPO could lead to more innovation activities by giving access to equity financing.

On the other hand, corporate finance literature widely documents that agency problems weaken the efficiency of a firm’s business operations (e.g., investment and M&A) after an IPO (Berle and Means, 1932; Jensen and Meckling, 1976). Similar analysis also applies to innovation activities. Innovation is often risky and could cost a manager his or her job. Career concerns may thus force a manager to choose less risky and therefore less innovative projects, which would eventually weaken the firm’s innovation activity. In addition, innovation is a difficult and time-consuming task that may require more time and effort. An entrenched manager, therefore, could put less effort into innovation and enjoy a “quieter lifestyle,” given the weaker institutional investor monitoring after the IPO. These two incentive channels would lead to less innovation.

With all these trade-offs, whether an IPO will increase or decrease a firm’s innovation remains an empirical question. Using uniquely matched Chinese firm-level data, our paper aims to address this dilemma. To that end, we use data on annual patent applications from China’s State Intellectual Property Office (SIPO) to capture firm innovation activities.

First, we use the number of patents that firms apply for each year to measure the quantity of innovation.

Second, we measure the quality of innovation by utilizing the Chinese patent categorization system. China’s patent system divides patents into three categories: invention, design, and utility. Invention patents are considered major innovations, and go through rigorous and lengthy examination of substance similar to that of the United States (U.S.). Design and utility patents are viewed as minor innovations, thereby receiving much looser examination of substance and weaker protection compared to invention patents.² Prior literature, which mainly focuses on United States and European Union (EU) patent data, uses the patent citation number as a proxy for the quality of innovation (e.g., He and Tian, 2013). As citation data are not available for Chinese patents, we take advantage of the unique characteristics of the Chinese patent system and use the number of invention patent applications (alternatively the proportion of invention patents and the number of inventors per patent) to measure patent quality.

Last, the detailed information available on the patents’ technology class allows us to explore a firm’s innovation scope. We match patent application data with Chinese Industrial Survey (CIS) data, which contain detailed balance sheet information for all industrial firms (both listed and non-listed) that are either state-owned or non-state-owned with annual sales greater than RMB 5 million (around USD 800,000) (See Section 3.1 for details on the matching of the two datasets). Finally, we use the Chinese Stock Market & Accounting Research Database (CSMAR) to identify an IPO firm and its year of IPO.

Having these uniquely matched IPO datasets, we first run an ordinary least squares (OLS) regression using panel data as the baseline. Our results suggest that IPOs are positively associated with firm innovation, both quantity and quality. To be more specific, public listing is associated with an average 35.1 percent annual increase in total patent applications in the subsequent three years. A similar increase also holds for invention patent applications, indicating that IPOs are associated with an increase in quantity as well as quality.

An important concern in OLS analysis is that public listing could be endogenous, thus subjecting results to selection bias. For example, some factors could be simultaneously correlated with the firm’s IPO decision and its innovation activity. As Jain and Kini (1994) point out, firms choose to go public at a specific stage in their life cycle, which could correspond to a certain level of firm innovation activities.

To mitigate these endogeneity concerns, we employ a difference-in-differences (DiD) approach that compares the innovation output of IPO firms or IPO subsidiaries (i.e. subsidiaries that were established before the IPO and continued to be subsidiaries during the three years following the IPO) to that of non-IPO firms and subsidiaries before and after the IPO conversion. To validate the DiD approach and ensure that IPO firms are compared to similar unlisted firms, we carefully match each IPO firm (and subsidiary) in the treatment group with a firm in the same industry but outside the treatment group by using the propensity score matching (PSM) algorithm along a broad range of firm characteristics. After undertaking a series of diagnostic tests to ensure the removal of observable differences between the two groups, we find that firms in the treatment group generate an average increase of 22.9 percent in annual patent applications following an IPO. The quality of the patent, as indicated by invention patent applications, appears to also increase by 12.9 percent. Other measurements of quality, such as the invention patent ratio and the number of inventors per patent, similarly indicate a rate increase in quality of innovation after the IPO.

To strengthen the identification process, we further use the two-step Heckman-type endogenous switching regression method to identify the causal relationship between IPO and innovation. This test generates quantitative results for innovation quantity and quality that are very similar to those of the DiD test, which corroborates our baseline findings.

We further examine the change in scope of innovation in the DiD framework by dividing a firm’s patent applications into two categories—related and unrelated—based on whether a patent’s technology area is related to the firm’s core business or not. Theoretically, capital raised from an IPO could be used either to help a firm strengthen its competitive advantage in its core business or to facilitate its expansion into a new business. Our empirical evidence shows that IPO firms see an increase in innovation activity at both their core and new business.

The DiD estimation approach helps us mitigate the endogeneity problem in the OLS regression and shows that IPOs do indeed have a positive impact on innovation activity. To gain a better understanding of the factors that cause such positive effect on innovation, we further apply a difference-in-difference-in-differences (DiDiD) framework that helps determine how the IPO impact varies across firms’ financial constraints, corporate governance, and ownership structures.

We further demonstrate that financial constraints play an important role in determining innovation performance after an IPO by addressing a firm’s financial constraints from two perspectives.

First, we test whether the impact of an IPO on innovation depends on a firm’s external financing needs. Following Rajan and Zingales (1998), we divide all two-digit industries into two groups: industries dependent on external finance (those above the median level of the external finance dependence index) and industries able to draw on internal finance (those below the median level of the external finance dependence index). Higher external finance needs imply that a firm is more likely to be subject to financial constraints. Our DiDiD regressions indicate that firms in industries dependent on external finance tend to experience higher increase rates in innovation (in terms of both quantity and quality) after an IPO compared to firms in industries with access to internal finance. This finding points to financial constraints playing an important role in a firm’s increase in innovation after an IPO. The equity raised by an IPO helps relax a firm’s financial constraints and enables it to meet its financing needs, which leads, in turn, to an increase in innovation activity.

Second, we look at the impact of collateral constraints on innovation. Following the literature (e.g., Kiyotaki and Moore, 1997), we use the ratio of fixed assets to total assets as a proxy for collateral constraints. A high ratio implies a higher collateral value; hence, less stringent financial constraints. We find evidence to support that an IPO firm with a lower fixed-assets ratio (characterized by stricter collateral constraints) before an IPO tends to experience improved innovation after the IPO. This finding is consistent with the message from the test of external financing needs: the equity raised by an IPO helps relax financial constraints, thus leading to increasing innovation activity. These results help us better understand the positive impact of IPOs on firm innovation in China. With an under-developed financial system, firms in China often face severe financial constraints. Equity fundraising through IPOs offers them a valuable source of innovation and growth.

In addition to financial constraints, our results show that corporate governance structures also play an important role in the positive impact of IPOs on innovation. We find that IPOs have a stronger impact on firms granting stock to their managers. By aligning the interests of managers and shareholders, stock granting provides a stronger incentive to innovate so as to add long-term value for the firm. Moreover, we find that firms with the same person as CEO and president tend to innovate more after an IPO. Career Concern Theory argues that, given the high risk of failure associated with innovation, managers may avoid engaging in it not to risk their own job security. According to this theory, CEOs who are also the president of their firms are entrenched (likely to be founders or even controlling shareholders) and thus more willing to take risk. Our results lend support to the Career Concern Theory.

Moreover, we discovered that privately-owned enterprises (POEs) exhibit a greater increase in innovation after an IPO than state-owned enterprises (SOEs). Since POEs face more stringent financial constraints before an IPO and deal better with agency problems caused by IPOs than SOEs, this finding further corroborates the importance of alleviating financial constraints as well as the deep effect agency problems have on innovation activity after an IPO.

Given the substantive increase in firm innovation quantity and quality, and the expansion of innovation scope after an IPO, it is natural to wonder how it happens. Leveraging the inventor information in our patent data, we show that having more inventors and better retention of existing inventors might be the reasons underlying the enhanced innovation activity after a firm’s IPO. Our DiD test indicates that IPO firms tend to hire more inventors than non-IPO firms. This implies that the accumulation of human capital played a role in the increased number of patent applications. In addition, our analysis of inventors further demonstrates that IPO firms (especially POEs) have much lower inventor departure rates than non-IPO firms, indicating that IPOs help firms retain their innovators. Finally, we do not find evidence that the increase in patent applications is due to a growth in the productivity of inventors after an IPO. Therefore, enhanced innovation seems to be associated with changes in the extensive margin (i.e., entry and exit) rather than the intensive margin (productivity per existing inventor).

Further, we explore the following bottom-line question: Does the enhanced innovation activity after an IPO help create firm value in the long run? Our answer is yes. This question is important because IPO firms may simply increase their patent applications for “window-dressing” purposes. We have two pieces of evidence against the “window-dressing” hypothesis: First, as mentioned above, the quality of the patent applications (measured by invention patent counts, the share of invention patents in total patents, and the number of inventors for each patent) increases after an IPO. Second, we find that firms that apply for more patents during the three years following an IPO see a larger Tobin’s Q increase, indicating that patent applications are associated with long-term value creation. Moreover, we find that the increase in invention patents has a stronger impact on increasing Tobin’s Q than other patent types. All the above support the argument that increased innovation activity after an IPO leads to firm value creation in the long run and disprove that window dressing may boost short-term valuation.³

The rest of the paper is organized as follows: Section 2 discusses the related literature. Section 3 describes the data sample and variable construction; it also reports summary statistics. Section 4 presents baseline results and addresses identification issues concerning innovation quantity, quality, and scope. Section 5 further explores the relationships among IPOs, inventors, and innovation activity. Section 6 investigates the impact of enhanced innovation after an IPO on a firm’s value. Section 7 provides the robustness check of the causal relationship between IPOs and innovation through exploring the two-step Heckman-type endogenous switching regression. Section 8 concludes.

II. Review of the Existing Literature

Our paper contributes to three strands of literature. First, studies on the real impact of IPOs have documented a wide range of post-IPO firm performance changes, such as the decline in profitability—documented in Jain and Kini (1994); Pagano, Panetta, and Zingales (1998); and Pastor, Taylor, and Veronesi (2009)—and the reduction in productivity shown in Chemmanur, He, and Nandy (2009). This paper contributes to the literature by leveraging our data and proposing a new DiD strategy to estimate the IPO effect on firms’ innovative activities in the Chinese context.

Additionally, the paper is closely related to contemporaneous research by Bernstein (2015), who examines a similar question in the context of US IPOs. Our analysis differs from Bernstein’s (2015) in several ways: We have implemented a DiD methodology to tackle the identification problem to leverage our unique datasets, including the balance sheet information of both listed and non-listed manufacturing firms in China. In contrast, Bernstein (2015) draws on an instrumental variable methodology using US IPO firm data.⁴

Additionally, we find that an IPO can substantially enhance firm innovation activity in terms of quantity, quality, and scope, which is in sharp contrast to the findings of Bernstein (2015) stating that IPOs have no impact on firm innovation quantity and have a significantly negative impact on quality. Our paper complements Bernstein (2015) by using a different identification strategy and emphasizing the positive impact of IPOs on innovation in the Chinese context, where firms face different financial constraints and corporate governance structures compared to their counterparts in the U.S. Our findings suggest that IPO benefits outweigh the costs of firm innovation in China and, perhaps even in other comparable developing countries. Moreover, we believe that financial constraints might play a very important role in understanding the reason behind our findings. Firms in China often face much more severe financial constraints than firms in the U.S., which makes accessing the equity market of uppermost importance for innovation in China. This aspect explains our finding that Chinese firms have more financial resources for retaining inventors after an IPO. Although we also identify the existence of the agency problem in China, which according to Bernstein (2015) is the source of the major negative impact on innovation brought by IPOs in the U.S., the benefits of IPOs in relaxing financial constraints significantly outweigh the negative impact of the agency problem induced by IPOs, a factor that had not been explored in the prior literature.

Second, the paper is also related to publications on equity markets and innovation, including a growing strand of literature comparing the behavior of public and private firms along various dimensions such as investment sensitivity (e.g., Asker, Farre-Mensa, and Ljungqvist, 2010; Sheen, 2009), debt financing and borrowing costs (Saunders and Steffen, 2011; Brav, 2009), external financing needs (Acharya and Xu, 2016), dividend payouts (Michaely and Roberts, 2007), and CEO compensation (Gao, Lemmon, and Li, 2010). Atanassov, Nanda, and Seru (2007) argue that arm’s length financing (equity and public debt) is positively related to innovation, while relationship-based bank financing is negatively related to innovation. Our paper advances this line of inquiry by providing new evidence showing the positive role that arm’s length financing can play in innovation.

Third, this paper also contributes to a growing theoretical and empirical literature strand that explores the role of governance, capital structure, and ownership concentration in corporate innovation. For instance, literature on larger institutional ownership (discussed by Aghion, Van Reenen, and Zingales, 2013), corporate venture capital (Chemmanur et al., 2014), and hotrather-than-cold markets (Nanda and Rhodes-Kropf, 2013) all treat how altered managerial incentives motivate managers to focus more on long-term innovation activities. Our paper provides new evidence showing that career concerns and management incentives can play an important role in affecting firm innovation after IPOs.

III. Data Description

3.2 Variable Construction

We use the patent filing counts to quantify innovation activity. The distribution of patent applications in our final sample is right-skewed, similar to U.S. patent data, with its median at zero (See Table 1). Following the literature (e.g., He and Tian, 2013; Gu, Mao, and Tian, 2013), we winsorize these variables at the 99^th percentile. We then use the natural logarithm of one plus patent counts, LnPat, as the main measure of innovation activity. As illustrated previously, since invention patents are widely regarded as being of higher quality compared to design and utility patents, we also use the natural logarithm of one plus invention patent applications (Lnlnv) as the quality measure of firm innovation.

Table 1.

Summary statistics

This table presents summary statistics for variables constructed using our merged data sample of listed and non-listed firms from 1998 to 2007. All variables are winsorized at the 1^st and 99^th percentiles of their distribution. The variable abbreviations are explained in the Table A1 of the Appendix.

Table 1.

Summary statistics

This table presents summary statistics for variables constructed using our merged data sample of listed and non-listed firms from 1998 to 2007. All variables are winsorized at the 1^st and 99^th percentiles of their distribution. The variable abbreviations are explained in the Table A1 of the Appendix.

	N	Mean	S.D.	P25	Median	P75
Pat	999114	0.18	1.91	0.00	0.00	0.00
Invention	999114	0.03	0.62	0.00	0.00	0.00
Ln(AT)	999114	10.04	1.37	9.06	9.87	10.87
FixAT_AT	999114	0.33	0.20	0.17	0.30	0.45
Interest_AT	999114	0.01	0.02	0.00	0.01	0.02
Admin_AT	999114	0.08	0.08	0.03	0.06	0.10
Ln(Age)	999114	2.22	0.80	1.79	2.20	2.71
Leverage	999114	0.59	0.27	0.40	0.60	0.78
Liquidity	999114	0.06	0.29	−0.11	0.06	0.24
ROA	999114	0.08	0.15	0.00	0.03	0.10
HI	999114	0.05	0.07	0.01	0.02	0.05
HI2	999114	0.01	0.05	0.00	0.00	0.00
SOE	999114	0.12	0.33	0.00	0.00	0.00
EX	999114	0.34	0.47	0.00	0.00	1.00
Ln(Stock)	999114	0.18	0.58	0.00	0.00	0.00

View Table

Table 1.

Summary statistics

This table presents summary statistics for variables constructed using our merged data sample of listed and non-listed firms from 1998 to 2007. All variables are winsorized at the 1^st and 99^th percentiles of their distribution. The variable abbreviations are explained in the Table A1 of the Appendix.

	N	Mean	S.D.	P25	Median	P75
Pat	999114	0.18	1.91	0.00	0.00	0.00
Invention	999114	0.03	0.62	0.00	0.00	0.00
Ln(AT)	999114	10.04	1.37	9.06	9.87	10.87
FixAT_AT	999114	0.33	0.20	0.17	0.30	0.45
Interest_AT	999114	0.01	0.02	0.00	0.01	0.02
Admin_AT	999114	0.08	0.08	0.03	0.06	0.10
Ln(Age)	999114	2.22	0.80	1.79	2.20	2.71
Leverage	999114	0.59	0.27	0.40	0.60	0.78
Liquidity	999114	0.06	0.29	−0.11	0.06	0.24
ROA	999114	0.08	0.15	0.00	0.03	0.10
HI	999114	0.05	0.07	0.01	0.02	0.05
HI2	999114	0.01	0.05	0.00	0.00	0.00
SOE	999114	0.12	0.33	0.00	0.00	0.00
EX	999114	0.34	0.47	0.00	0.00	1.00
Ln(Stock)	999114	0.18	0.58	0.00	0.00	0.00

In our regression analysis, we follow the innovation literature in controlling for a set of firm and industry characteristics that might affect a firm’s future innovation output. In the baseline regression, the control variables are (i) firm size, measured by the natural logarithm of total assets, Ln(AT); (ii) FixAT_AT, measured by net fixed assets divided by total assets; (iii) Admin_AT, measured by administration expenditure divided by total assets; (iv) Interest_AT, measured by interest expenditure divided by total assets; (v) Leverage, measured by total liabilities divided by total assets; (vi) Liquidity, measured by the ratio of the difference between current assets and liabilities and total assets; (vii) ROA, measured by net income divided by total assets; (viii) Ln(Age), measured by the natural logarithm of one plus the number of years after establishment; (ix) HI, measured by the Herfindahl index based on annual sales; (x) SOE, a state-owned enterprise dummy that takes the value of one for SOEs, zero otherwise; (xi) EX, an exporter dummy that is equal to one for exporters, zero otherwise; and (xii) Stock, patent stock measured by the total patent applications of the firm prior to the current year. To circumvent the potential nonlinear effects of product market competition (Aghion et al, 2005), we include the squared Herfindahl index in our baseline regressions. All variables are computed for firm i over its fiscal year t.

IV. Empirical Results

4.1 Time Series Pattern and Panel Regression Results

In Figure 1, we plot the average number of industry- and year-adjusted innovation output variables within a seven-year window around the IPO event. More specifically, we adjust total patent (invention patent) filing numbers by subtracting the average number of total patent (invention patent) filings of all firms in the same industry and year. The time series plots in both Panel A and Panel B demonstrate that total patent (including invention, utility, and design patents) and invention patent filings do not significantly increase over the three years leading up to firms’ IPOs. In contrast, there is sharp jump in the number of both total patent filings and invention patent filings starting from the IPO year. The increase lasts through the three years following the IPO.

Industry and year-adjusted innovation output around IPOs

This figure presents firms’ average number of industry- and year-adjusted patent applications around the IPO year. We adjust each firm’s patent applications by subtracting the average number of patent applications of all firms in the same two-digit industry and year. Panel A shows that the time trend for total patent counts (including invention, design, and utility patents) and Panel B shows the time trend for invention patent counts only. The x-axis indicates the number of years before or after the IPO.

Citation: IMF Working Papers 2017, 147; 10.5089/9781484303726.001.A001

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Industry and year-adjusted innovation output around IPOs

This figure presents firms’ average number of industry- and year-adjusted patent applications around the IPO year. We adjust each firm’s patent applications by subtracting the average number of patent applications of all firms in the same two-digit industry and year. Panel A shows that the time trend for total patent counts (including invention, design, and utility patents) and Panel B shows the time trend for invention patent counts only. The x-axis indicates the number of years before or after the IPO.

Citation: IMF Working Papers 2017, 147; 10.5089/9781484303726.001.A001

• Download Figure
• Download figure as PowerPoint slide

Industry and year-adjusted innovation output around IPOs

This figure presents firms’ average number of industry- and year-adjusted patent applications around the IPO year. We adjust each firm’s patent applications by subtracting the average number of patent applications of all firms in the same two-digit industry and year. Panel A shows that the time trend for total patent counts (including invention, design, and utility patents) and Panel B shows the time trend for invention patent counts only. The x-axis indicates the number of years before or after the IPO.

Citation: IMF Working Papers 2017, 147; 10.5089/9781484303726.001.A001

• Download Figure
• Download figure as PowerPoint slide

Inspired by the raw pattern of innovation output surrounding the IPO in the time series data as demonstrated in Figure 1, we use the IPO firms’ firm-year observations for a seven-year window centered on the IPO event and estimate an OLS regression:

$\begin{array}{cc}Pa{t}_{it}\left(in{v}_{it}\right)=\alpha +{\beta }_{1}befor{e}_{it}^{-3}+{\beta }_{2}befor{e}_{it}^{-2}+{\beta }_{3}curren{t}_{it}+{\beta }_4{after}_{it}^1+{\beta }_5{after}_{it}^2+{\beta }_6{after}_{it}^3+{ϵ}_{it}& \left(1\right)\end{array}$

where i denotes the firm and t denotes the year. The dependent variable is firm i’s total patent filing (invention patent filing) adjusted for the industry-year average in year t. The independent variable ${before}_{it}^{-3}$ is a dummy that equals one if the firm-year observation is three years before the IPO event, and zero otherwise. Similarly, ${before}_{it}^{-2}$ is a dummy that equals one if the firm-year observation is two years before the IPO event, and zero otherwise. current_it is a dummy that equals one if the firm-year observation is in the IPO year, and zero otherwise. The dummy variables ${after}_{it}^1$, ${after}_{it}^2$, ${after}_{it}^3$ are equal to one when the observations are one year, two years, and three years after the IPO event respectively, and zero otherwise. As can be seen, the benchmark group comprises the observations one year before the IPO.

The results are reported in Table 2. We find that the coefficient estimates of ${before}_{it}^{-2}$ and ${before}_{it}^{-3}$ are statistically insignificant for both total patents and invention patents, showing that firm innovation output stays at the same level before the IPO. In contrast, the coefficients of current, ${after}_{it}^1$, ${after}_{it}^2$ and ${after}_{it}^3$ are all positive and statistically significant, indicating a jump in innovation activity, both in terms of quantity and quality, after the IPO. These findings confirm the time series trend shown in Panels A and B of Figure 1.

Next, to take advantage of our rich data set, we explore the impact of IPOs on firm innovation using a panel data regression approach. We estimate the following regression

$\begin{array}{cc}{Lnpat}_{i,t+n}\left({LnIn}_{i,t+n}\right)=\alpha +\beta {list}_{it}+\gamma X_{it}+{year}_t+{industry}_{ij}+{ϵ}_{it}& \left(2\right)\end{array}$

where LnPat_i,t+n(Lnlnv_u+n) is the natural logarithm of one plus the total patent (invention patent) filing in year t+n for firm i, where n is equal to 1,2,3. The variable of interest, list_it, is equal to one if firm i or firm i’s parent firm undergoes an IPO in year t, and zero otherwise. X_it is a group of firm-level control variables, summarized in Table 1, which according to the literature (e.g., Gu, Mao, and Tian 2013) are relevant to a firm’s innovation activities. year_t and industry_ij control for the year and industry-level fixed effects.

The results of the panel data regression are reported in Table 2 Panel B. The coefficient estimate of list_it is positive and statistically significant at the 1 percent level for n=1,2,3. Based on the coefficients estimated in columns (1)-(3), going public is associated with an increase in average annual patent filings of 35.1 percent for the subsequent three years after the IPO, indicating an economically significant magnitude. Similar results also hold for the invention patent applications according to columns (4)-(6). Various robustness tests were conducted to corroborate our baseline OLS regression results. For instance, we employed quantile regressions with various specifications, and the coefficients for the list dummy variable are always positive and significant.

Table 2 Panel A:

Raw patterns for innovation dynamics surrounding IPOs

This table reports the OLS regression results estimating the innovation dynamics surrounding IPOs. Using a sample of all listed firms, we retain firm-year observations for a seven-year window centered on the IPO year, and we estimate the pooled OLS regression of the following model:

$Pa{t}_{it}\left(In{v}_{it}\right)=\alpha +{\beta }_{1}befor{e}_{it}^{-3}+{\beta }_{2}befor{e}_{it}^{-2}+{\beta }_{2}curren{t}_{it}+{\beta }_4{after}_{it}^1+{\beta }_{5}afte{r}_{it}^2+{\beta }_{6}afte{r}_{it}^3+{ϵ}_{it}$

The dependent variable is either Pat_it, firm i’s industry- and year-adjusted total patents filed in year t, or Inv_it, firm i ‘ s industry- and year-adjusted invention patents filed in year t. We adjust innovation output variables by subtracting the average values of innovation outputs for all firms (excluding the listed firms) in the same two-digit industry and year. ${before}_{it}^{-3}\left({before}_{it}^{-2}\right)$ is a dummy that equals one if a firm-year observation is from 3 (2) years before the IPO year, and zero otherwise. current_it is a dummy that equals one if a firm-year observation is in its IPO year, and zero otherwise. ${after}_{it}^1({after}_{it}^2,{after}_{it}^3)$ is a dummy that equals one if a firm-year observation is 1 (2, 3) year after the IPO year, and zero otherwise. Therefore, the omitted group (benchmark) comprises the observations 1 year before the IPO year. Standard errors clustered by industry are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

Table 2 Panel A:

Raw patterns for innovation dynamics surrounding IPOs

This table reports the OLS regression results estimating the innovation dynamics surrounding IPOs. Using a sample of all listed firms, we retain firm-year observations for a seven-year window centered on the IPO year, and we estimate the pooled OLS regression of the following model:

$Pa{t}_{it}\left(In{v}_{it}\right)=\alpha +{\beta }_{1}befor{e}_{it}^{-3}+{\beta }_{2}befor{e}_{it}^{-2}+{\beta }_{2}curren{t}_{it}+{\beta }_4{after}_{it}^1+{\beta }_{5}afte{r}_{it}^2+{\beta }_{6}afte{r}_{it}^3+{ϵ}_{it}$

The dependent variable is either Pat_it, firm i’s industry- and year-adjusted total patents filed in year t, or Inv_it, firm i ‘ s industry- and year-adjusted invention patents filed in year t. We adjust innovation output variables by subtracting the average values of innovation outputs for all firms (excluding the listed firms) in the same two-digit industry and year. ${before}_{it}^{-3}\left({before}_{it}^{-2}\right)$ is a dummy that equals one if a firm-year observation is from 3 (2) years before the IPO year, and zero otherwise. current_it is a dummy that equals one if a firm-year observation is in its IPO year, and zero otherwise. ${after}_{it}^1({after}_{it}^2,{after}_{it}^3)$ is a dummy that equals one if a firm-year observation is 1 (2, 3) year after the IPO year, and zero otherwise. Therefore, the omitted group (benchmark) comprises the observations 1 year before the IPO year. Standard errors clustered by industry are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

VARIABLES	(1) Pat	(2) Inv
before3	−0.262 (0.313)	−0.086 (0.102)
before2	−0.054 (0.218)	−0.045 (0.069)
current	0.433* (0.220)	0.246* (0.123)
after1	1.012** (0.423)	0.610*** (0.212)
after2	0.757** (0.292)	0.561*** (0.197)
after3	1.255** (0.532)	0.820** (0.315)
Constant	1.465*** (0.369)	0.302*** (0.061)
Observations	2,118	2,118
R-squared	0.005	0.012

View Table

Table 2 Panel A:

Raw patterns for innovation dynamics surrounding IPOs

This table reports the OLS regression results estimating the innovation dynamics surrounding IPOs. Using a sample of all listed firms, we retain firm-year observations for a seven-year window centered on the IPO year, and we estimate the pooled OLS regression of the following model:

$Pa{t}_{it}\left(In{v}_{it}\right)=\alpha +{\beta }_{1}befor{e}_{it}^{-3}+{\beta }_{2}befor{e}_{it}^{-2}+{\beta }_{2}curren{t}_{it}+{\beta }_4{after}_{it}^1+{\beta }_{5}afte{r}_{it}^2+{\beta }_{6}afte{r}_{it}^3+{ϵ}_{it}$

The dependent variable is either Pat_it, firm i’s industry- and year-adjusted total patents filed in year t, or Inv_it, firm i ‘ s industry- and year-adjusted invention patents filed in year t. We adjust innovation output variables by subtracting the average values of innovation outputs for all firms (excluding the listed firms) in the same two-digit industry and year. ${before}_{it}^{-3}\left({before}_{it}^{-2}\right)$ is a dummy that equals one if a firm-year observation is from 3 (2) years before the IPO year, and zero otherwise. current_it is a dummy that equals one if a firm-year observation is in its IPO year, and zero otherwise. ${after}_{it}^1({after}_{it}^2,{after}_{it}^3)$ is a dummy that equals one if a firm-year observation is 1 (2, 3) year after the IPO year, and zero otherwise. Therefore, the omitted group (benchmark) comprises the observations 1 year before the IPO year. Standard errors clustered by industry are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

VARIABLES	(1) Pat	(2) Inv
before3	−0.262 (0.313)	−0.086 (0.102)
before2	−0.054 (0.218)	−0.045 (0.069)
current	0.433* (0.220)	0.246* (0.123)
after1	1.012** (0.423)	0.610*** (0.212)
after2	0.757** (0.292)	0.561*** (0.197)
after3	1.255** (0.532)	0.820** (0.315)
Constant	1.465*** (0.369)	0.302*** (0.061)
Observations	2,118	2,118
R-squared	0.005	0.012

Table 2 Panel B:

OLS regression of innovation outcomes on IPOs

This table reports the OLS results estimating the effect of an IPO on innovation output variables. We estimate the pooled OLS regression of the following model:

${LnPat}_{it+n}\left({LnInv}_{it+n}\right)=\alpha +{\beta list}_{it}+{\gamma X}_{it}+{year}_t+{industry}_{ij}+{ϵ}_{it}$

using a sample of all listed and non-listed firms from 1998 to 2007. The dependent variable LnPat_it+n is the natural logarithm of one plus the total number of patents applied for one (t+1), two (t+2), and three (t+3) years after year t, and the results are reported in columns (1)–(3), respectively. The dependent variable Lnlnv_it+n is the natural logarithm of one plus the number of invention patents applied for one (t+1), two (t+2), and three (t+3) years after year t, and the results are reported in columns (4)-(6), respectively. List_it is a dummy variable that equals one if an IPO occurs in year t for firm i, and zero otherwise. Year fixed effects, year_it, and industry fixed effects, industry_it, are included in all regressions. All other variables are as defined in the Appendix. Standard errors clustered by industry are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

Standard errors in parentheses*** p<0.01, ** p<0.05, *p<0.1

Table 2 Panel B:

OLS regression of innovation outcomes on IPOs

This table reports the OLS results estimating the effect of an IPO on innovation output variables. We estimate the pooled OLS regression of the following model:

${LnPat}_{it+n}\left({LnInv}_{it+n}\right)=\alpha +{\beta list}_{it}+{\gamma X}_{it}+{year}_t+{industry}_{ij}+{ϵ}_{it}$

using a sample of all listed and non-listed firms from 1998 to 2007. The dependent variable LnPat_it+n is the natural logarithm of one plus the total number of patents applied for one (t+1), two (t+2), and three (t+3) years after year t, and the results are reported in columns (1)–(3), respectively. The dependent variable Lnlnv_it+n is the natural logarithm of one plus the number of invention patents applied for one (t+1), two (t+2), and three (t+3) years after year t, and the results are reported in columns (4)-(6), respectively. List_it is a dummy variable that equals one if an IPO occurs in year t for firm i, and zero otherwise. Year fixed effects, year_it, and industry fixed effects, industry_it, are included in all regressions. All other variables are as defined in the Appendix. Standard errors clustered by industry are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

VARIABLES	(1) LnPat_+1	(2) LnPat_+2	(3) LnPat_+3	(4) Lnlnv_+1	(5) Lnlnv_+2	(6) Lnlnv_+3
List	0.336*** (0.071)	0.312*** (0.053)	0.406*** (0.093)	0.205*** (0.041)	0.183*** (0.038)	0.263*** (0.076)
Ln(AT)	0.048*** (0.007)	0.050*** (0.007)	0.052*** (0.008)	0.016*** (0.003)	0.017*** (0.003)	0.019*** (0.003)
FixAT_AT	−0.066*** (0.013)	−0.071*** (0.014)	−0.079*** (0.015)	−0.017*** (0.004)	−0.022*** (0.005)	−0.025*** (0.006)
Admin_a	0.214*** (0.035)	0.223*** (0.037)	0.235*** (0.039)	0.062*** (0.014)	0.066*** (0.015)	0.074*** (0.016)
Ln(Age)	−0.004** (0.002)	−0.005*** (0.002)	−0.005*** (0.002)	−0.001 (0.001)	−0.001 (0.001)	−0.001 (0.001)
SOE	0.008 (0.006)	0.009 (0.006)	0.010 (0.006)	0.006** (0.003)	0.006** (0.003)	0.006** (0.003)
EX	0.025*** (0.006)	0.026*** (0.007)	0.026*** (0.007)	0.006*** (0.002)	0.007*** (0.002)	0.007** (0.003)
Leverage	−0.024*** (0.004)	−0.023*** (0.004)	−0.024*** (0.005)	−0.007*** (0.002)	−0.009*** (0.002)	−0.009*** (0.002)
Liquidity	−0.007* (0.004)	−0.007* (0.004)	−0.008* (0.004)	0.001 (0.001)	−0.000 (0.002)	−0.001 (0.002)
Interest_At	−0.033 (0.032)	−0.056* (0.033)	−0.047 (0.037)	0.036** (0.017)	0.030 (0.019)	0.029 (0.019)
ROA	0.053*** (0.009)	0.068*** (0.011)	0.088*** (0.015)	0.020*** (0.004)	0.028*** (0.006)	0.038*** (0.008)
HI	−0.034 (0.026)	−0.037 (0.027)	−0.050* (0.029)	−0.003 (0.011)	−0.006 (0.013)	−0.009 (0.013)
HI2	0.066* (0.034)	0.076** (0.036)	0.081** (0.038)	0.023 (0.020)	0.029 (0.023)	0.028 (0.025)
Constant	−0.440*** (0.065)	−0.459*** (0.069)	−0.482*** (0.072)	−0.159*** (0.028)	.0.171*** (0.031)	−0.186*** (0.032)
Observations	839,761	709,997	580,425	839,761	709,997	580,425
R-squared	0.064	0.065	0.067	0.039	0.040	0.043
Year fixed effects	Y	Y	Y	Y	Y	Y
Industry fixed effect	Y	Y	Y	Y	Y	Y
Firm fixed effects	N	N	N	N	N	N

Standard errors in parentheses*** p<0.01, ** p<0.05, *p<0.1

View Table

Table 2 Panel B:

OLS regression of innovation outcomes on IPOs

This table reports the OLS results estimating the effect of an IPO on innovation output variables. We estimate the pooled OLS regression of the following model:

${LnPat}_{it+n}\left({LnInv}_{it+n}\right)=\alpha +{\beta list}_{it}+{\gamma X}_{it}+{year}_t+{industry}_{ij}+{ϵ}_{it}$

using a sample of all listed and non-listed firms from 1998 to 2007. The dependent variable LnPat_it+n is the natural logarithm of one plus the total number of patents applied for one (t+1), two (t+2), and three (t+3) years after year t, and the results are reported in columns (1)–(3), respectively. The dependent variable Lnlnv_it+n is the natural logarithm of one plus the number of invention patents applied for one (t+1), two (t+2), and three (t+3) years after year t, and the results are reported in columns (4)-(6), respectively. List_it is a dummy variable that equals one if an IPO occurs in year t for firm i, and zero otherwise. Year fixed effects, year_it, and industry fixed effects, industry_it, are included in all regressions. All other variables are as defined in the Appendix. Standard errors clustered by industry are reported in parentheses. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

VARIABLES	(1) LnPat_+1	(2) LnPat_+2	(3) LnPat_+3	(4) Lnlnv_+1	(5) Lnlnv_+2	(6) Lnlnv_+3
List	0.336*** (0.071)	0.312*** (0.053)	0.406*** (0.093)	0.205*** (0.041)	0.183*** (0.038)	0.263*** (0.076)
Ln(AT)	0.048*** (0.007)	0.050*** (0.007)	0.052*** (0.008)	0.016*** (0.003)	0.017*** (0.003)	0.019*** (0.003)
FixAT_AT	−0.066*** (0.013)	−0.071*** (0.014)	−0.079*** (0.015)	−0.017*** (0.004)	−0.022*** (0.005)	−0.025*** (0.006)
Admin_a	0.214*** (0.035)	0.223*** (0.037)	0.235*** (0.039)	0.062*** (0.014)	0.066*** (0.015)	0.074*** (0.016)
Ln(Age)	−0.004** (0.002)	−0.005*** (0.002)	−0.005*** (0.002)	−0.001 (0.001)	−0.001 (0.001)	−0.001 (0.001)
SOE	0.008 (0.006)	0.009 (0.006)	0.010 (0.006)	0.006** (0.003)	0.006** (0.003)	0.006** (0.003)
EX	0.025*** (0.006)	0.026*** (0.007)	0.026*** (0.007)	0.006*** (0.002)	0.007*** (0.002)	0.007** (0.003)
Leverage	−0.024*** (0.004)	−0.023*** (0.004)	−0.024*** (0.005)	−0.007*** (0.002)	−0.009*** (0.002)	−0.009*** (0.002)
Liquidity	−0.007* (0.004)	−0.007* (0.004)	−0.008* (0.004)	0.001 (0.001)	−0.000 (0.002)	−0.001 (0.002)
Interest_At	−0.033 (0.032)	−0.056* (0.033)	−0.047 (0.037)	0.036** (0.017)	0.030 (0.019)	0.029 (0.019)
ROA	0.053*** (0.009)	0.068*** (0.011)	0.088*** (0.015)	0.020*** (0.004)	0.028*** (0.006)	0.038*** (0.008)
HI	−0.034 (0.026)	−0.037 (0.027)	−0.050* (0.029)	−0.003 (0.011)	−0.006 (0.013)	−0.009 (0.013)
HI2	0.066* (0.034)	0.076** (0.036)	0.081** (0.038)	0.023 (0.020)	0.029 (0.023)	0.028 (0.025)
Constant	−0.440*** (0.065)	−0.459*** (0.069)	−0.482*** (0.072)	−0.159*** (0.028)	.0.171*** (0.031)	−0.186*** (0.032)
Observations	839,761	709,997	580,425	839,761	709,997	580,425
R-squared	0.064	0.065	0.067	0.039	0.040	0.043
Year fixed effects	Y	Y	Y	Y	Y	Y
Industry fixed effect	Y	Y	Y	Y	Y	Y
Firm fixed effects	N	N	N	N	N	N

Standard errors in parentheses*** p<0.01, ** p<0.05, *p<0.1

4.2 Difference-in-differences Test

A legitimate concern regarding panel data regression results is the endogeneity problem. It is possible that both the IPO decision and firm innovation are affected by firms’ unobservable characteristics. For instance, a rapidly growing firm is more likely to go public, and at the same time, it has reached a growth stage whereby it wants to increase its innovation activity. Our identification strategy to alleviate the endogeneity concern is based on a difference-in-differences (DiD) approach. The CIS data, which provide detailed balance sheet information for non-publicly listed manufacturing firms, allow us to match the IPO firms with non-publicly listed firms that have similar firm characteristics to the IPO firms. We then compare the innovation output of the IPO firms to the innovation output of the matching firms, which do not go public but are otherwise comparable, before and after the IPO.

The DiD approach has important advantages. First, it can rule out omitted trends that are correlated with an IPO and innovation in both treatment and control groups. For instance, the prosperity of a certain industry can simultaneously increase the likelihood of an IPO and future innovation. The DiD approach rules out the possibility that industry prosperity, rather than the IPO itself, drives the change in innovation activity. Second, the DiD approach controls for constant unobservable differences between the treatment and control groups.

We construct the treatment and control groups of firms using the PSM algorithm. Specifically, the treatment group includes all IPO firms and their subsidiary firms for the period 1999–2006 with non-missing matching variables in the pre-IPO year (t-1) and the post-IPO year (t+1). We further require that the subsidiary firms must maintain their subsidiary status from IPO year t to year t+3 and they must be established no later than t-3.

We use the PSM algorithm to identify matches between IPO firms and firms that do not go public. In our PSM, we first estimate a probit model with 356 firms associated with IPOs and 772,734 firm-year observations without IPOs. The dependent variable Y_it equals one if firm i has an IPO at year t, and zero otherwise. In the probit model, we include all control variables from the baseline specification in equation (2), which are measured in the year immediately preceding the IPO. We also include patent growth during the three years before the IPO, g_pn, and the patent stock, Stock, in the probit model, to ensure that the parallel trend assumption—the key requirement for the DiD approach—is appropriately accommodated.

The probit regression results indicate that our model captures a significant amount of variation in the dependent variable. We report the probit model results in column (1) of Panel A in Table 3. As can be seen, the pseudo R² reaches 16.5 percent and the p-value from the Chi-squared test of the overall model fitness is well below 0.0001, indicating that our model specification works well. We then use the predicted probabilities, or propensity scores, to conduct a nearest-neighbor PSM procedure. Specifically, we match each IPO firm-year observation (i.e., treatment group) with a non-IPO firm-year observation that has the closest propensity score among the observations in that IPO year and belongs to the same industry as the IPO firm to form the control group. We eventually generate 355 pairs of matched firms.⁷

Table 3.

Difference-in-differences (DiD) test results

Table 3.

Difference-in-differences (DiD) test results

This table reports diagnostic tests and the DiD results concerning how IPOs affect firm innovation. Our sample contains firms that experienced an IPO from 1999 to 2006 and non-listed firms. We match firms using a one-to-one nearest neighbor PSM, with replacement, on a host of observable characteristics, including all independent variables used in equation (2), for the year before the IPO, the growth in the number of patents, g_pn, computed over the three-year period before the IPO, two-digit industry dummies, and year fixed effects. Definitions of all other variables are listed in the Appendix. Our treatment group contains those listed firms’ observations in their IPO year. Our control group includes all non-listed firm-year observations. Panel A reports parameter estimates from the probit model used in estimating the propensity scores for the treatment and control groups. The dependent variable equals one for the firm-year belonging to the treatment group and zero for those belonging to the control group. The “Pre-Match” column contains the parameter estimates of the probit model estimated using the sample prior to matching. These estimates are then used to generate the propensity scores for matching. The “Post-Match” column contains the parameter estimates of the probit model estimated using the subsample of matched treatment-control pairs after matching. Robust standard errors are displayed in parentheses below each coefficient estimate. Panel B presents the univariate comparisons between the treatment and control firms’ characteristics, and their corresponding p-values, testing the null hypothesis that the differences are zero. Panel C gives the DiD test results. Patent is the mean of firm i’s number of patents in the three-year window before or after the IPO. Invention is the average of firm i’s number of inventions in the three-year window before or after the IPO. Panel D reports the regression results estimating the innovation dynamics of the treatment and control firms surrounding an IPO. The dependent variable is either Ln Patit, the natural logarithm of one plus firm i’s number of patents in year t, or Ln Invit, the natural logarithm of one plus firm i’s number of inventions in year t. listi is a dummy that equals one for treatment firms (listed firms) and zero for control firms (non-listed firms). ${before}_{it}^{2\&3}$ is a dummy that equals one if a firm-year observation is from 2 or 3 years before the IPO year (year 0), and zero otherwise. currentit is a dummy that equals one if a firm-year observation is in the IPO year (year 0), and zero otherwise. ${after}_{it}^1({after}_{it}^2,{after}_{it}^3)$ is a dummy that equals one if a firm-year observation is from 1 (2,3) year after the IPO year (year 0), and zero otherwise. Robust standard errors are displayed in parentheses under each coefficient estimate. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

View Table

Table 3.

Difference-in-differences (DiD) test results

This table reports diagnostic tests and the DiD results concerning how IPOs affect firm innovation. Our sample contains firms that experienced an IPO from 1999 to 2006 and non-listed firms. We match firms using a one-to-one nearest neighbor PSM, with replacement, on a host of observable characteristics, including all independent variables used in equation (2), for the year before the IPO, the growth in the number of patents, g_pn, computed over the three-year period before the IPO, two-digit industry dummies, and year fixed effects. Definitions of all other variables are listed in the Appendix. Our treatment group contains those listed firms’ observations in their IPO year. Our control group includes all non-listed firm-year observations. Panel A reports parameter estimates from the probit model used in estimating the propensity scores for the treatment and control groups. The dependent variable equals one for the firm-year belonging to the treatment group and zero for those belonging to the control group. The “Pre-Match” column contains the parameter estimates of the probit model estimated using the sample prior to matching. These estimates are then used to generate the propensity scores for matching. The “Post-Match” column contains the parameter estimates of the probit model estimated using the subsample of matched treatment-control pairs after matching. Robust standard errors are displayed in parentheses below each coefficient estimate. Panel B presents the univariate comparisons between the treatment and control firms’ characteristics, and their corresponding p-values, testing the null hypothesis that the differences are zero. Panel C gives the DiD test results. Patent is the mean of firm i’s number of patents in the three-year window before or after the IPO. Invention is the average of firm i’s number of inventions in the three-year window before or after the IPO. Panel D reports the regression results estimating the innovation dynamics of the treatment and control firms surrounding an IPO. The dependent variable is either Ln Patit, the natural logarithm of one plus firm i’s number of patents in year t, or Ln Invit, the natural logarithm of one plus firm i’s number of inventions in year t. listi is a dummy that equals one for treatment firms (listed firms) and zero for control firms (non-listed firms). ${before}_{it}^{2\&3}$ is a dummy that equals one if a firm-year observation is from 2 or 3 years before the IPO year (year 0), and zero otherwise. currentit is a dummy that equals one if a firm-year observation is in the IPO year (year 0), and zero otherwise. ${after}_{it}^1({after}_{it}^2,{after}_{it}^3)$ is a dummy that equals one if a firm-year observation is from 1 (2,3) year after the IPO year (year 0), and zero otherwise. Robust standard errors are displayed in parentheses under each coefficient estimate. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

Table 3 Panel A:

Pre-match propensity score regression and post-match diagnostic regression

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Table 3 Panel A:

Pre-match propensity score regression and post-match diagnostic regression

VARIABLES	(1) Pre-match	(2) Post-match
Ln(AT)	0.256*** (0.013)	−0.046 (0.043)
FixAT_AT	−0.373*** (0.100)	0.045 (0.365)
Admin_AT	0.863*** (0.178)	0.141 (0.992)
Ln(Age)	−0.185*** (0.021)	0.000 (0.054)
SOE	0.243*** (0.043)	0.061 (0.112)
EX	−0.081** (0.036)	−0.007 (0.113)
Leverage	−0.284*** (0.086)	0.048 (0.333)
Liquidity	−0.006 (0.096)	−0.074 (0.298)
Interns_AT	0.915 (0.980)	1.645 (3.962)
ROA	0.644*** (0.086)	0.499 (0.510)
HI	0.151 (0.420)	0.125 (1.265)
HI2	−0.599 (0.778)	−1.046 (2.248)
Stock	0.098*** (0.019)	0.077 (0.050)
g_pn	0.050 (0.030)	0.045 (0.077)
Constant	−5441*** (0.183)	0.408 (0.627)
Observations	773,090	710
P-value of Chi-squared	0.0000	1.0000
Pseudo-R2	0.1652	0.0066
Industry fixed effects	Y	Y
Year fixed effects	Y	Y

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

View Table

Table 3 Panel A:

Pre-match propensity score regression and post-match diagnostic regression

VARIABLES	(1) Pre-match	(2) Post-match
Ln(AT)	0.256*** (0.013)	−0.046 (0.043)
FixAT_AT	−0.373*** (0.100)	0.045 (0.365)
Admin_AT	0.863*** (0.178)	0.141 (0.992)
Ln(Age)	−0.185*** (0.021)	0.000 (0.054)
SOE	0.243*** (0.043)	0.061 (0.112)
EX	−0.081** (0.036)	−0.007 (0.113)
Leverage	−0.284*** (0.086)	0.048 (0.333)
Liquidity	−0.006 (0.096)	−0.074 (0.298)
Interns_AT	0.915 (0.980)	1.645 (3.962)
ROA	0.644*** (0.086)	0.499 (0.510)
HI	0.151 (0.420)	0.125 (1.265)
HI2	−0.599 (0.778)	−1.046 (2.248)
Stock	0.098*** (0.019)	0.077 (0.050)
g_pn	0.050 (0.030)	0.045 (0.077)
Constant	−5441*** (0.183)	0.408 (0.627)
Observations	773,090	710
P-value of Chi-squared	0.0000	1.0000
Pseudo-R2	0.1652	0.0066
Industry fixed effects	Y	Y
Year fixed effects	Y	Y

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Before progressing to the DiD test, we conduct a number of diagnostic tests to check the validity of the parallel trend assumption, which is essential for the DiD approach. First, we rerun the probit model, restricting the observations to only the 710 matched firms after the pairing. The results are shown in column (2) of Panel A in Table 3. As can be seen, none of the coefficients of the independent variables are statistically significant. In particular, the coefficient of the pre-IPO patent application growth variable g_pm is insignificant, indicating that there is no observable difference in innovation trend between the IPO and non-IPO firms before the IPO event. Moreover, the pseudo R² falls dramatically from 16.52 percent to 0.66 percent after the matching and the p-value of the Chi-squared test is close to 1, indicating that the null hypothesis that all the coefficient estimates are zero cannot be rejected.

Furthermore, we conduct a series of univariate variable tests of firm characteristics between the treatment and control group firms one year before the IPO. The results are reported in Panel B of Table 3, with the p-value of the tests shown in the last column of the table. As can be seen, none of the differences in firm characteristics between the treatment and control groups are statistically significant after the matching. In particular, the insignificance of the pre-IPO patent application growth variable (g_pn) provides further evidence that the parallel trend assumption is likely to hold.

Table 3 Panel B:

Post-match differences

Table 3 Panel B:

Post-match differences

	Control	Treat	Treat-Control	p-value
Ln(AT)	11.94	11.86	−0.08	0.45
FixAT_AT	0.29	0.29	0.00	0.85
Admin_AT	0.06	0.07	0.00	0.34
Ln(Age)	1.99	2.01	0.01	0.86
SOE	0.37	0.37	0.01	0.88
EX	0.42	0.42	0.00	1.00
Leverage	0.53	0.53	0.00	0.87
Liquidity	0.11	0.11	0.00	0.81
Interest_AT	0.01	0.01	0.00	0.63
ROA	0.08	0.09	0.01	0.36
HI	0.06	0.06	0.00	0.67
HI2	0.01	0.01	0.00	0.53
Stock	0.57	0.69	0.12	0.16
g_pn	0.12	0.14	0.04	0.40

View Table

Table 3 Panel B:

Post-match differences

	Control	Treat	Treat-Control	p-value
Ln(AT)	11.94	11.86	−0.08	0.45
FixAT_AT	0.29	0.29	0.00	0.85
Admin_AT	0.06	0.07	0.00	0.34
Ln(Age)	1.99	2.01	0.01	0.86
SOE	0.37	0.37	0.01	0.88
EX	0.42	0.42	0.00	1.00
Leverage	0.53	0.53	0.00	0.87
Liquidity	0.11	0.11	0.00	0.81
Interest_AT	0.01	0.01	0.00	0.63
ROA	0.08	0.09	0.01	0.36
HI	0.06	0.06	0.00	0.67
HI2	0.01	0.01	0.00	0.53
Stock	0.57	0.69	0.12	0.16
g_pn	0.12	0.14	0.04	0.40

In summary, the above diagnostic tests provide strong evidence that our PSM process removes significant observable differences (other than the difference in IPO) between the treatment and control group firms, meaning that the change in innovation activity is more likely to be driven by IPOs in the DiD test.

To provide a visual demonstration of the trend in patent applications around the IPO for both the treatment and control group firms, we plot the time series of the variable LnPat (LnInv) during a seven-year window around the IPO event in Figure 2. As can be seen, the trend lines of the treatment and control groups move closely in parallel during the years leading up to the IPO event, which provides another piece of evidence for the parallel trend assumption. However, after the IPO event, we can see a significant increase in patent applications for the treatment group line. In sharp contrast, the control group line stays flat after the IPO.

Innovation outputs around IPOs for treatment and control firms

This figure presents the mean values of innovation outputs around the IPO event. Panel A shows the time trend for the natural logarithm of one plus patent counts and Panel B shows the time trend for the natural logarithm of one plus invention counts. The square line is for treatment firms that implemented an IPO in a sample year. The diamond line is for PSM non-list firms (control group). The x-axis indicates the years before or after the IPO.

Citation: IMF Working Papers 2017, 147; 10.5089/9781484303726.001.A001

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Innovation outputs around IPOs for treatment and control firms

This figure presents the mean values of innovation outputs around the IPO event. Panel A shows the time trend for the natural logarithm of one plus patent counts and Panel B shows the time trend for the natural logarithm of one plus invention counts. The square line is for treatment firms that implemented an IPO in a sample year. The diamond line is for PSM non-list firms (control group). The x-axis indicates the years before or after the IPO.

Citation: IMF Working Papers 2017, 147; 10.5089/9781484303726.001.A001

• Download Figure
• Download figure as PowerPoint slide

Innovation outputs around IPOs for treatment and control firms

This figure presents the mean values of innovation outputs around the IPO event. Panel A shows the time trend for the natural logarithm of one plus patent counts and Panel B shows the time trend for the natural logarithm of one plus invention counts. The square line is for treatment firms that implemented an IPO in a sample year. The diamond line is for PSM non-list firms (control group). The x-axis indicates the years before or after the IPO.

Citation: IMF Working Papers 2017, 147; 10.5089/9781484303726.001.A001

• Download Figure
• Download figure as PowerPoint slide

To corroborate the visual evidence, we conduct formal DiD tests and report the results in Panels C and D of Table 3. Panel C shows the change in average annual total patent (invention patent) applications around an IPO. Specifically, the first column in Panel C reports the difference in average annual total patent (invention patent) applications during the three years before and after the IPO event for the treatment group. Similarly, we calculate the change in average annual total patent (invention patent) applications for the control group in the second column. These results indicate that the changes in both total patent and invention patent applications are positive and statistically significant for the treatment group. The changes, however, are insignificant for the control group. The third and fourth columns further report the DiD estimate results for the treatment and control groups, respectively. We find that the DiD tests for the total patent and invention patent filings are both positive and statistically significant at the 1 percent level. The DiD test results are not only statistically significant, but are also economically significant. On average, an IPO results in an increase of 1.56 (0.77) total patent (invention patent) applications annually over the three years after the IPO.

Table 3 Panel C:

DID estimates

Table 3 Panel C:

DID estimates

	Mean Treatment Difference (after-before)	Mean Control Difference (after-before)	Mean DID estimate (treat-control)	p-value for DID Estimate
Patent	1.577	0.014	1.563	0.001
(s.e.)	0.405	0.262	0.482
Inv	0.858	0.086	0.772	0.000
(s.e.)	0.198	0.072	0.210

View Table

Table 3 Panel C:

DID estimates

	Mean Treatment Difference (after-before)	Mean Control Difference (after-before)	Mean DID estimate (treat-control)	p-value for DID Estimate
Patent	1.577	0.014	1.563	0.001
(s.e.)	0.405	0.262	0.482
Inv	0.858	0.086	0.772	0.000
(s.e.)	0.198	0.072	0.210

Table 3 Panel D:

DiD analysis for innovation dynamics

Standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Table 3 Panel D:

DiD analysis for innovation dynamics

VARIABLES	(1) LnPat	(2) Lnlnv
${list}_i*{before}_{it}^{2\&3}$	−0.043 (0.052)	−0.066 (0.041)
listt * current	0.075 (0.050)	0.043 (0.031)
${list}_i*{after}_{it}^1$	0.164*** (0.052)	0.093*** (0.033)
${list}_i*{after}_{it}^2$	0.231*** (0.054)	0.101*** (0.034)
${list}_i*{after}_{it}^3$	0.291*** (0.062)	0.192*** (0.042)
${before}_{it}^2$	0.080** (0.038)	0.043* (0.023)
${before}_{it}^1$	0.116*** (0.044)	0.062*** (0.024)
Current	0.096** (0.042)	0.047* (0.024)
${after}_{it}^1$	0.096** (0.043)	0.068*** (0.024)
${after}_{it}^2$	0.060 (0.047)	0.071** (0.029)
${after}_{it}^3$	0.127** (0.050)	0.078*** (0.028)
Constant	0.173*** (0.044)	0.064** (0.025)
Observations	4,253	4,253
R-squared	0.655	0.576
Year fixed effects	Y	Y
Firm fixed effects	Y	Y

Standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

View Table

Table 3 Panel D:

DiD analysis for innovation dynamics

VARIABLES	(1) LnPat	(2) Lnlnv
${list}_i*{before}_{it}^{2\&3}$	−0.043 (0.052)	−0.066 (0.041)
listt * current	0.075 (0.050)	0.043 (0.031)
${list}_i*{after}_{it}^1$	0.164*** (0.052)	0.093*** (0.033)
${list}_i*{after}_{it}^2$	0.231*** (0.054)	0.101*** (0.034)
${list}_i*{after}_{it}^3$	0.291*** (0.062)	0.192*** (0.042)
${before}_{it}^2$	0.080** (0.038)	0.043* (0.023)
${before}_{it}^1$	0.116*** (0.044)	0.062*** (0.024)
Current	0.096** (0.042)	0.047* (0.024)
${after}_{it}^1$	0.096** (0.043)	0.068*** (0.024)
${after}_{it}^2$	0.060 (0.047)	0.071** (0.029)
${after}_{it}^3$	0.127** (0.050)	0.078*** (0.028)
Constant	0.173*** (0.044)	0.064** (0.025)
Observations	4,253	4,253
R-squared	0.655	0.576
Year fixed effects	Y	Y
Firm fixed effects	Y	Y

Standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Panel D in Table 3 further demonstrates the innovation dynamics of the DiD results in a regression framework. The firm-year observations are gathered in a seven-year window centered on the IPO year. The regression model is as follows:

$\begin{array}{cc}InPa{t}_{it}\left(InIn{v}_{it}\right)=\alpha +{\beta }_{1}lis{t}_i*befor{e}_{it}^{-2\&-3}+{\beta }_{2}lis{t}_i*curren{t}_{it}+{\beta }_{3}lis{t}_i*afte{r}_{it}^1+{{\beta }_{4}lis{t}_i*{after}_{it}^2+\beta }_{5}lis{t}_i*afte{r}_{it}^3+befor{e}_{it}^{-2}+befor{e}_{it}^{-1}+curren{t}_{it}+afte{r}_{it}^1+afte{r}_{it}^2+afte{r}_{it}^3+yea{r}_t+fir{m}_i+{ϵ}_{it}& \left(3\right)\end{array}$

where i denotes the firm and t denotes the time; list_t is a dummy, equal to one if firm i belongs to the treatment group (IPO firms), and zero otherwise (non-IPO firms); ${before}_{it}^{-2\&-3}$, current_it, ${after}_{it}^1$, ${after}_{it}^2$ and afterf_t are all dummy variables equal to one if the year t is three or two years before an IPO, the IPO year, and one year, two years, or three years after the IPO, respectively, and zero otherwise. Therefore, the omitted or benchmark group comprises the observations one year before the IPO. We also include the year and firm fixed effects in the regression.

The coefficient estimates of interest are β₁,β₂,β₃,β₄,β₅. If reverse causality holds, that is, the increase in innovation activities helps improve the quality of the firm, which then leads to an IPO, we should observe significant and positive coefficient estimates of β₁. However, in contrast, we find statistically insignificant and negative coefficient estimates of β₁, which indicates that there is no systemically different trend in firm innovation activity between IPO firms and non-IPO firms before the IPO. The coefficient estimates of β₃,β₄, and β₅ are all positive and statistically significant, suggesting that compared to the non-IPO firms, IPO firms experience a significant increase in their patent application activity after the IPO event. Overall, our findings suggest that IPOs help increase firm innovation activity and reverse causality does not hold.

To check the robustness of our results, we also repeat the PSM algorithm and construct the control group samples by choosing five (instead of one) most closely matched non-IPO firms for each IPO firm in the treatment group. In our unreported results, the DiD tests on the treatment group and the augmented control group generate quantitatively similar results as in Table 3.