How Do Government Subsidies Affect Innovation? Evidence from Chinese Hi-Tech SMEs

Abstract

This paper examines the effectiveness of government subsidies in fostering innovation among small and medium-sized enterprises (SMEs), with a particular focus on additionality, crowding-out, and cherry-picking effects. Using the latest national survey data on Chinese high-tech SMEs, we apply robust econometric techniques—including the Heckman selection model, structural equation modeling (SEM), and propensity score matching (PSM)—to address potential selection bias and endogeneity. Our findings reveal that government subsidies positively influence both innovation inputs and outputs, suggesting a predominant additionality effect rather than a crowding-out effect, at least within high-tech SMEs. However, subsidies do not appear to alleviate the financial constraints faced by most SMEs, indicating that they are insufficient as a standalone solution to financing challenges. Furthermore, state ownership enhances input additionality but does not significantly impact output additionality. We also find evidence of cherry-picking in subsidy allocation, with loans exhibiting stronger additionality effects on innovation compared to grants and tax credits, which are more prone to selective intervention. These findings highlight the need for more targeted subsidy policies that prioritize financially constrained firms with high innovation potential while mitigating government selectivity. Our study offers valuable insights for policymakers seeking to design more effective innovation support mechanisms for high-tech SMEs.

1. Introduction

The effectiveness of government subsidies in stimulating innovation has long been debated, both in private industry [1,2] and in SMEs specifically [3,4]. Recent studies suggest that public support for private innovation can reduce unit costs and enhance the expected profitability of funded innovation projects [5]. Government subsidies also help alleviate financial constraints and incentivize firms to invest in innovation, thereby stimulating additional innovation activities [6,7]—a process known as additionality. However, as Zúñiga-Vicente et al. [8] argue, government subsidies may also lead firms to substitute public funds for their own R&D investment, merely crowding out private spending. Additionally, governments or their agencies may misuse subsidy programs to serve their own interests [9]. For example, the selection of firms for innovation funding often prioritizes those with a high likelihood of success, rather than those where public aid would have the greatest marginal impact. This cherry-picking approach creates an illusion of policy effectiveness, falsely crediting both the funding bodies and the subsidized programs while overlooking more deserving projects [10]. Such inefficiencies are particularly pronounced in emerging economies with weaker institutional and legal frameworks, where officials responsible for subsidy allocation may exploit opaque distribution mechanisms for personal gain [11].

Beyond the direct relationship between government subsidies and corporate innovation, scholars have explored additional dimensions of this policy tool. Wang et al. [12] examined the spatial effects of government subsidies on corporate innovation but found no evidence of positive spillover effects among firms. Wang et al. [13] analyzed the influence of political connections and investor attention on the effectiveness of subsidies, revealing that political connections suppress innovation incentives, while investor attention enhances them. Furthermore, Xu et al. [14] empirically demonstrated that tax incentives more effectively improve green innovation efficiency in non-state-owned enterprises, as well as firms in their growth and maturity stages. These findings highlight the need for more precise and context-specific subsidy policies, tailored to different economic environments and stages of enterprise development, to maximize their intended benefits.

These nuanced findings regarding subsidy effectiveness across institutional contexts and firm characteristics call for examination within specific policy ecosystems where innovation financing mechanisms are most actively evolving. China presents a compelling case study, as its multi-layered subsidy architecture—combining centralized R&D directives with decentralized implementation—has created unique dynamics for high-tech SMEs. Notably, over 50% of national innovation budgets are executed at subnational levels through specialized mechanisms like stage investments and risk subsidies, making Chinese local governments crucial innovation policy laboratories.

The research focus on the role of government subsidies in Chinese hi-tech SMEs is motivated by the fact that in China economic development through innovation, high technology (hi-tech), and business R&D has substantially been progressing along with financial support from the public toward the private sector. At the national level, the Ministry of Science and Technology (MOST) administers major R&D programmes such as the key technologies programme (KTP) and the national high-tech programme (NHTP). Furthermore, Chinese local government and agencies have also recently become important funders of innovation in their jurisdictions [15]. In fact, more than half of Chinese spending on support for science and technology, including R&D programmes, is at the subnational level. This clearly shows the increasing importance of local government in implementing innovation policy. These funds are further extended to promote the innovation activities of SMEs [16]. As a result of this policy, regional and local government in China, along with the China Development Bank, fund hi-tech SMEs and their innovation. These funds are directly allocated in one of these three forms of investment: (i) stage investment; (ii) risk subsidies; and (iii) investment protection. (Stage investment refers to funds invested in hi-tech SMEs as non-controlling shareholders (less than 25% of total equity). Under risk subsidies, the venture investment institutions are entitled to claim a certain percentage of the risk compensation funds from the risk subsidies in the event of project losses. Lastly, investment protection refers to pre-investment and post-investment protection for the proposed business by the venture investment institutions, such that R&D expenditures and market-oriented innovation are subsidized.) In contrast to these forms of direct support, tax credits and credit guarantees are the most common form of indirect support for innovation in Chinese SMEs.

This paper provides a fresh analysis that is intended to contribute to recent studies into the role of government subsidies in the following ways. First, in dealing with innovation measurement, we employ both innovation inputs (as measured by R&D intensity) and innovation outputs (as measured by sales based on innovation). Dimos and Pugh [17] show that a positive effect may cancel out a full crowding-out effect with respect to innovation inputs and outputs. Therefore, we estimate both the positive and negative effects of government subsidization on both innovation inputs and outputs. In particular, we argue that a positive effect of government subsidies on innovation can demonstrate the greater likelihood of innovation additionality for SMEs. That also allows us to examine innovation efficiency by comparing innovation inputs and outputs. Second, we address the complicating endogeneity issue that often arises in these sorts of studies. The problem is how to control for endogeneity in the models that are associated with selection bias and causality. To this end, we deploy the Heckman approach to deal with the selection problem. In the next step we then use a simultaneous equation model (SEM) to cope with the endogeneity arising from the allocation of government subsidies, i.e., the cherry-picking effect. To deal with the challenge to discriminate between the potential different effects induced by heterogeneous treatment levels, provided to firms through the subsidies, and to isolate the effects of the subsidies from other confounding factors [1], we take into account the different effects of three common types of government subsidies, comprising direct support in the form of grants and loans and indirect, represented by tax credits. These types of support differ in design, timing, cost and effect. For example, Montmartin and Herrera [2] argue that direct support implies relatively heavier administrative costs, while indirect support favours projects with high private returns and direct support seems to be linked to projects with considerable social returns. Finally, empirical studies critically depend on the selection of the reliable dataset. We therefore use a unique data set on Chinese hi-tech SMEs collected from the latest national survey sponsored by the National Social Science Fund of China (NSSFC) in 2015. This dataset enables us to focus on the effect of government subsidies on hi-tech SMEs, a sector particularly targeted for innovation policy.

Our empirical findings indicate that government subsidies positively relate to innovation input and output. Considering the financial constraints faced by most SMEs, we conclude that government subsidies are thus more likely to have additionality than crowding-out effects when it comes to innovation, at least in hi-tech SMEs. Moreover, state ownership has a positive effect on input additionality but not output additionality. We also find that both innovation inputs and outputs have a two-way relationship with government subsidies, thereby providing evidence of the practice of cherry picking in subsidy allocation. In particular, we find that the three types of subsidies exhibit a significant difference in the effects. After comparing the signs and magnitudes of the estimated coefficients, we argue that loans invoke relatively more innovation inputs and outputs through the additionality effect. Both grants and tax credits are affected by cherry picking, where grants act like “a bonus”; tax credits are affected by a post refund.

The remainder of the paper is structured as follows. Section 2 develops the hypotheses used in the paper. Section 3 discusses the data, variables and estimation models. Section 4 presents and discusses the results. Section 5 sets out the conclusion.

2. Literature Review and Hypotheses

2.1. Subsidies, Crowding out and Additionality

Government intervention in innovation is fundamentally justified by market failure theory, which identifies three critical barriers to efficient private investment: (1) information asymmetry between innovators and financiers that obscures project valuation [18], (2) the long-term horizon and multi-stage risks inherent to innovation activities [19], and (3) knowledge spillovers that erode private returns [20]. These factors collectively result in suboptimal innovation funding in capital markets [18], a problem disproportionately affecting small and medium enterprises (SMEs).

The academic debate on subsidy efficacy centers on two opposing mechanisms. The additionality perspective posits that public subsidies stimulate new R&D investments that would otherwise not occur, particularly for resource-constrained firms [7,21]. Conversely, the crowding-out view argues subsidies merely displace private R&D expenditures, as firms treat public funds as substitutes for other financing [17,22]. Empirical ambiguity persists due to contextual variations across national systems [11], industry characteristics [23], subsidy instruments [24], and methodological approaches [2,25].

SMEs face amplified challenges due to structural constraints: their informational opacity [26], limited access to formal financing [27,28], and tendency toward novel, high-risk innovations [29] exacerbate market failures. This leads to chronic R&D underinvestment [30] and heightened reliance on public support [31]. While subsidies can alleviate these constraints [32], their effectiveness hinges on balancing additionality gains against crowding-out risks. Drawing on those perspectives, we formulate the following hypothesis:

Hypothesis 1.

Government subsidies provide greater additionality than crowding-out effects on SME innovation.

2.2. Subsidies and State Ownership

Since the 1980s, China’s Reform and Opening policy has driven significant ownership reforms in state-owned enterprises (SOEs), aimed at enhancing efficiency and competitiveness. The 1995 policy of “keeping large firms and letting small ones go” marked a strategic shift: large SOEs in key sectors (e.g., energy, utilities) remained under state control, while smaller SOEs underwent privatization through debt-equity swaps, ownership diversification, and management buyouts [33]. Over decades, this reform transformed most state-owned SMEs into joint-stock entities, alongside the rise of private SMEs. However, despite reduced state ownership in the SME sector compared to large firms, many SMEs remain state-influenced, particularly through local governments that control subsidy allocations.

State ownership confers distinct advantages on SMEs, including preferential access to capital, monopoly profits in strategic industries, and policy-driven financial support [34,35,36]. State-owned banks prioritize lending to SOEs to meet broader policy goals like employment stability [37], while political connections through government-controlled associations facilitate innovation financing [38]. Local governments, as both owners and subsidy allocators, further amplify these benefits, enabling state-influenced SMEs to overcome financing constraints and secure R&D resources [39,40]. These institutional advantages suggest that state ownership may enhance the effectiveness of subsidies in boosting innovation inputs. We therefore formulate the following hypothesis:

Hypothesis 2a.

State ownership has positive and moderate effects on the relationship between government subsidies and SME innovation input.

Nevertheless, state ownership introduces inefficiencies that undermine innovation outcomes [41]. Soft budget constraints [42] reduce pressure for efficient resource use, while bureaucratic governance stifles strategic innovation investments [43]. Even minority state ownership may distort subsidy utilization, as seen in the misallocation of R&D funds [25]. Furthermore, state-dominated firms often exhibit lower efficiency in converting resources into innovation [39], partly due to rigid oversight and misaligned incentives [44]. While state backing eases financial constraints, it may simultaneously suppress innovation output by prioritizing stability over risk-taking. Based on these views, we may formulate the next hypothesis:

Hypothesis 2b.

State ownership has negative and moderate effects on the relationship between government subsidies and SME innovation output.

2.3. Subsidies and Cherry Picking

Cherry-picking in subsidy allocation differs fundamentally from additionality and crowding-out effects, as it originates from the strategic behavior of subsidy providers rather than firm responses. Rooted in public choice theory, this phenomenon arises when policymakers prioritize projects with visible success metrics to maximize political capital and bureaucratic budgets [17]. In emerging economies with underdeveloped legal frameworks, such behavior is exacerbated by weak accountability mechanisms, leading officials to favor low-risk, high-visibility projects over socially optimal investments [11]. This misallocation not only diverts subsidies from high-potential but riskier innovations [10] but also creates perverse incentives: applicants may strategically conceal critical information to exploit subsidized funding as a low-cost financing tool, particularly when institutional oversight is limited [45]. The resulting information asymmetry perpetuates a cycle where bureaucrats secure short-term political gains, while firms prioritize subsidy capture over genuine innovation.

The distortions caused by cherry-picking extend across the innovation process. At the input stage, firms may inflate R&D expenditures to signal eligibility, while at the output stage, they may overstate incremental improvements as breakthroughs. Notably, output-based cherry-picking is more prevalent due to the higher political salience of measurable results (e.g., patents, sales). This dual distortion undermines subsidy effectiveness, as funds flow to ex-post observable outcomes rather than ex-ante transformative ideas. Drawing on these perspectives, we formulate the following hypotheses:

Hypothesis 3a.

Cherry picking occurs with respect to innovation inputs, that is, more innovation inputs induce the greater likelihood of obtaining subsidies.

Hypothesis 3b.

Cherry picking occurs with respect to innovation outputs, that is, more innovation outputs induce the greater likelihood of obtaining subsidies.

Hypothesis 3c.

Cherry picking is more salient for innovation outputs than for innovation inputs.

3. Data and Methodology

3.1. Data

In our analysis, we deploy data from the “China Hi-Tech Small and Medium-sized Enterprises Dynamic Growth Survey” (CTSMEDGS). The CSMEDGS is a major project of the National Social Science Fund of China (NSSFC) that was conducted in 2015. The main objective was to build up a credit system and database for financial evaluation, credit rating and policy-related research for hi-tech SMEs. The CSMEDGS survey comprises basic businees information, firm external growth environments and firm internal growth factors. The effectiveness and reliability of the survey is supported by the fact that four Chinese universities administered 10,000 questionnaires for the survey in 2015 through onsite distribution and collection. The firms sampled by the China Statistics Bureau and the State Administration for Industry and Commerce were randomly taken from the literally millions of registered non-financial hi-tech firms in China that employ fewer than 1000 employees. (High-tech enterprises are recognized by the High-tech Enterprise Office in accordance with the “Administrative Measures for the Recognition of High-tech Enterprises” revised and issued by the state in 2016 in China.)

Our analysis begins with 2683 high-tech firms from the CSMEDGS dataset, applying rigorous screening criteria to ensure data reliability: firms with negative/zero net assets (106 excluded to remove insolvent entities), zero annual cash flow (352 excluded indicating operational abnormalities), financial firms (209 excluded due to distinct regulatory environments), and those with excessive missing data (key variables, 831 excluded). The final sample of 1185 non-financial SMEs reflects the inherent limitations of survey-based data, where response quality and completeness constrain sample size. To address potential selection bias, we implemented methodological safeguards: the Heckman Two-Stage Model corrected for non-random attrition by modeling exclusion probabilities, Structural Equation Modeling (SEM) accounted for latent variables influencing both selection and outcomes, and Propensity Score Matching (PSM) balanced observable characteristics between included and excluded firms. Diagnostic tests further support generalizability, showing no significant differences in core financial ratios, comparable industry distribution, and consistent regional coverage. These measures collectively ensure robustness despite sample constraints typical of SME survey research.

Table 1. Definitions and descriptive statistics.

	Definition	Mean	Std. Dev.	Min.	Max.
RD_intensity	Ordinal dummy variables for the ratio of R&D expenditure to total sales: <1% (1), 1–3% (2), 3–5% (3), 5–10% (4), and >10% (5).	2.361	1.168	1.000	5.000
Innovative_sales(mil)	The sales on new products and/or new technology	27.835	37.449	0.000	1132.472
Finance_difficulty	Seven-point Likert scale from extremely easy to extremely hard	4.143	1.408	1.000	7.000
Grant_dummy	Dummy variable with one if firm the firm obtained government grants in the last year and zero otherwise	0.173	0.378	0.000	1.000
Loan_dummy	Dummy variable with one if firm the firm obtained discount loans in the last year and zero otherwise	0.122	0.328	0.000	1.000
Tax_dummy	Dummy variable with one if firm the firm obtained tax credits in the last year and zero otherwise	0.181	0.386	0.000	1.000
Grant	The amount of government grants obtained in the last year in logarithm	0.338	4.023	0.000	100.000
Loan	The amount of government loans obtained in the last year in logarithm	0.317	3.219	0.000	90.250
Tax credit	The amount of government grants obtained in the last year in logarithm	0.556	7.053	0.000	181.460
State	Dummy variable with one if firm is owned by the state and zero otherwise	0.329	0.470	0.000	1.000
Risk-preference	Dummy variable with one if the firm is not averse to high risk and zero otherwise	0.058	0.234	0.000	1.000
Firm_age	Number of years since firm establishment	8.207	6.864	0.000	67.000
Sale_growth	Ordinal dummy variables for the sales growth rates: <10% (1), 10–20% (2), 20–30% (3), 30–40% (4), and >40% (5).	1.970	0.842	1.000	5.000
Asset	Total assets in 10,000 RMB	5214.000	13,853.000	0.830	237,600.000
Export	Dummy variable with one if firm has exports and zero otherwise	0.316	0.465	0.000	1.000
Joint_RD	Dummy variable with one if the firm is largely relying on joint R&D and zero otherwise	0.209	0.407	0.000	1.000
Self_RD	Dummy variable with one if the firm is largely relying on self R&D and zero otherwise	0.267	0.442	0.000	1.000
Knowledge_supplier	Dummy variable with one if the knowledge is from supplier and zero otherwise	0.571	0.495	0.000	1.000
Knowledge_customer	Dummy variable with one if the knowledge is from customers and zero otherwise	0.537	0.498	0.000	1.000
Knowledge_rival	Dummy variable with one if the knowledge is from business rivals and zero otherwise	0.256	0.436	0.000	1.000
Knowledge_university	Dummy variable with one if the knowledge is from universities and zero otherwise	0.211	0.408	0.000	1.000
Knowledge_association	Dummy variable with one if the knowledge is from association and zero otherwise	0.251	0.434	0.000	1.000
Knowledge_advice	Dummy variable with one if the knowledge is from advisory companies and zero otherwise	0.127	0.333	0.000	1.000
Tech_employee	to what extent firms can get competent employees for R&D	4.895	1.354	1.000	7.000
Market_tech	to what extent the technology in the market is developing rapidly	4.474	1.742	1.000	7.000
Internal_finance	Dummy variable with one if firm relies on internal finance and zero otherwise	0.672	0.469	0.000	1.000
External_debt	Dummy variable with one if firm relies on external debt and zero otherwise	0.714	0.452	0.000	1.000
External_equity	Dummy variable with one if firm relies on external equity and zero otherwise	0.116	0.321	0.000	1.000
Collateral	Dummy variable with one if major loans are collateralized and zero otherwise	0.549	0.498	0.000	1.000
Reward	Dummy variable with one if the firm has go a national reward from the government for the firm’s product over the last three years and zero otherwise	0.199	0.399	0.000	1.000

shows data statistics and variable definitions.

3.2. Methods and Variables

We measure innovation across three dimensions—inputs, outputs, and process (innovation activities). In theory, a government subsidy can affect any or all of them. In the literature, a number of the studies, especially on additionality or crowding out, focus on the effects on the innovation inputs [17]. By contrast, a few studies consider output additionality and innovation behaviour additionality (innovation activity additionality) [23]. This can be because innovation expenditures are a common proxy for innovation and input additionality can be clearly measured. We define additionality as the extent to which public support stimulates new R&D activity [7]. The precise measurement of input additionality is then the additional firm-financed R&D spending over the amount of the subsidy [17].

This approach is difficult to apply to the measurement of output additionality and behavioural additionality (innovation activity additionality). As Dimos and Pugh [17] suggest, there are five possible outcomes of innovation inputs to compare the magnitude of subsidies and R&D expenditures. However, there are only three possibilities when we examine the effects of subsidies on innovation outputs. We cannot clearly distinguish how much of the outputs are attributable to government subsidies. For this reason, we choose to estimate the positive or negative effects of the government subsidy on innovation input and output. More precisely, a positive (negative) effect can cancel out the crowding-out effect with respect to innovation inputs (outputs). Furthermore, a significant and positive effect of a government subsidy on innovation could demonstrate the greater likelihood of innovation additionality for SMEs. SMEs suffer from financial constraints and therefore it is more likely for financial support or subsidy to induce additional innovation inputs. The reasoning also lies in the effects of subsidy on innovation output. That is, a positive subsidy effect on innovation outcomes demonstrates the effectiveness of the subsidy in the general sense. Thus, a positive and significant effect can imply that a subsidy is more likely to have additionality effects on innovation.

An important issue that also needs to be addressed to minimize bias in our findings is the endogeneity problem. Endogeneity may arise from two possible sources. The first is selection or sampling bias. Following Ghoul et al. [46] and Wooldridge [47], we address sampling bias by deploying the Heckman approach. The second source of endogeneity is the allocation of the government subsidy through cherry picking. To deal with this, we use a simultaneous equations model (SEM). We also isolate the effects of the subsidy from other effects by including several control variables in our models. The baseline model has the following form:

$$\begin{array}{ll}Subsidy\_d& ={\alpha }_0+{\alpha }_{1}State+{\alpha }_{2}Risk\_preference+{\alpha }_{3}Firm\_age\\ & +{\alpha }_{4}Sale\_growth+{\alpha }_{5}Asset+{\alpha }_{6}Export+{\alpha }_{7}Internal\_finance\\ & +{\alpha }_{8}External\_debt+{\alpha }_{9}External\_equity+{\alpha }_{10}Collateral+{\alpha }_{11}Reward+{\alpha }_{12}Industry+\epsilon \end{array}$$

(1)

where Subsidy is a binary dummy variable that indicates whether the firm received any government subsidies including grants, loans or tax credits in the survey year. State is a dummy variable with a value of one if the firm is owned by the government and zero otherwise. The control variables are mainly firm characteristics and types of financing. With respect to firm characteristics, Risk-preference represents whether the firm is not averse to high risk [48]; Firm_age, Sale_growth, Asset, Export, Industry represent firm age, sales growth, total assets and whether the firm has exports, and industry, respectively. With respect to types of firm financing, we consider Internal_finance, External_equity, External_debt. Collaterals are used to control for the major financing sources for SMEs. Lastly, we use Reward to control for the potential connection between firm production and government recognition, deriving from the question “whether the firm has received a national reward from the government for the firm’s product over the last three years?” Government subsidies reflect the preferential policies that encourage the hi-tech firms to carry out research and development and transformation of technological achievements. The policies also form the core independent intellectual property rights of enterprises, and carry out production and operation activities on this basis. Therefore, a reward by the government authorities serves as a token of the government’s recognition of the hi-tech firm’ achievement of technological performance, which is one of the critical criteria for granting government subsidies [11].

In a second stage, we specify Equation (2) to examine to what extent government subsidies can affect SME financing (difficulty of financing or financial constraints).

$$\begin{array}{ll}Finance\_difficulty& ={\alpha }_0+{\alpha }_{1}State+{\alpha }_{2}Risk\_preference+{\alpha }_{3}Firm\_age+{\alpha }_{4}Sale\_growth+{\alpha }_{5}Asset\\ & +{\alpha }_{6}Export+{\alpha }_{7}Internal\_finance+{\alpha }_{8}External\_debt+{\alpha }_{9}External\_equity\\ & +{\alpha }_{10}Collateral+{\alpha }_{11}Reward+{\alpha }_{12}Industry+{\alpha }_{13}Subsidy+{\alpha }_{14}Invmill+\epsilon \end{array}$$

(2)

where the variable Finance_difficulty measures the extent to which the firm faces the difficulties with its financing. We use for this measurement a seven-point scale (1—extremely easily; 7—extremely hard). We understand that this subjective measurement is not an optimal solution, however, the self-reporting of financial constraints or financing status is not unusual in the literature. For instance, Canepa and Stoneman [49] suggest that the lack of finance relates to the likelihood that hi-tech small firms in the UK report a project as delayed or abandoned. Invmill is the value of the inverse Mills’ ratio (λ) estimated from Equation (1).

Next, we estimate the effects of government subsidies on innovation using the following regression:

$$\begin{array}{ll}Innovation& ={\alpha }_0+{\alpha }_{1}State+{\alpha }_{2}Risk\_preference+{\alpha }_{3}Firm\_age+{\alpha }_{4}Sale\_growth+{\alpha }_{5}Asset\\ & +{\alpha }_{6}Export+{\alpha }_{7}Internal\_finance+{\alpha }_{8}External\_debt+{\alpha }_{9}External\_equity\\ & +{\alpha }_{10}Collateral+{\alpha }_{11}Joint\_RD+{\alpha }_{12}Self\_RD+{\alpha }_{13}Market\_tech+{\alpha }_{14}Tech\_employee\\ & +{\alpha }_{15}Knowledge\_advice+{\alpha }_{16}Knowledge\_supplier+{\alpha }_{17}Knowledge\_Customer\\ & +{\alpha }_{18}Knowledge\_rival+{\alpha }_{19}Knowledge\_university+{\alpha }_{20}Knowledge\_association\\ & +{\alpha }_{21}Industry+{\alpha }_{22}Subsidy+{\alpha }_{23}Interaction+{\alpha }_{24}Invmill+\epsilon \end{array}$$

(3)

where Innovation is innovation input (R&D intensity) or output (sales on innovation). R&D intensity is an ordinal variable for the ratio of R&D expenditure to total sales: <1% (1), 1–3% (2), 3–5% (3), 5–10% (4), and >10% (5). The measurement of R&D intensity as ordinal categories rather than a continuous variable is justified by two key considerations. First, both empirical and policy rationales support this approach. The classification thresholds align with government innovation subsidy criteria (e.g., the Regulations for High-Tech Enterprise Certification), ensuring compatibility with real-world policy implementation. Additionally, discretizing continuous ratios corrects severe right-skewed distributions and mitigates heteroscedasticity in OLS residuals. Second, survey limitations constrain measurement precision. SMEs typically report R&D ratios in intervals (e.g., “1–3%”) rather than exact values, precluding the construction of continuous variables. Moreover, self-reported data exhibit significant clustering at policy-relevant thresholds, violating the continuity assumption required for parametric tests. Innovation sales are the sales (in logarithms) of new products and/or technology for the year. We selected the logarithm of innovative product sales as our core output metric based on three key considerations. First, sales figures directly reflect market validation of innovation outcomes, thereby circumventing the “patent paradox” (where many patents never achieve commercialization). This approach aligns with the Oslo Manual’s principle of emphasizing market-tested indicators. Second, given China’s unique context, we observe significant variance in patent quality alongside prevalent “strategic patents” filed primarily to obtain subsidies. Our 2015 survey data verifies innovation sales through tax records, ensuring audited reliability. Third, the log-transformed innovation sales metric simultaneously captures both technological and business model innovations. This helps avoid truncation bias caused by SMEs’ low patenting propensity, while demonstrating statistically significant correlation with patent metrics at the 1% level. In conclusion, the innovation sales measure demonstrates superior effectiveness when these comprehensive factors are considered. Elsewhere, Interaction is an interaction term between State and Subsidy, with a set of variables used to control innovation, including Knowledge_supplier, Knowledge_customer, Knowledge_rival, Knowledge_advice, Knowledge_university and Knowledge_association, which implies the different sources of knowledge from suppliers, customers, business rivals, advisory services, universities (research institutions), and associations [50]. Other innovation variables that we include, are Joint_RD (relying on joint R&D), Self_RD (relying on independent R&D), Tech_employee (how hard the firm can get competent employees for R&D), Market_tech (to what extent the technology in the market is developing rapidly). The definitions of all other variables are identical to those in Equation (2).

In examining the cherry-picking hypothesis as well as dealing with the endogeneity arising from the allocation of government subsidies, we specify the simultaneous equation model (SEM) as follows:

$$\begin{array}{ll}Innovation& ={\alpha }_0+{\alpha }_{1}State+{\alpha }_{2}Risk\_preference+{\alpha }_{3}Firm\_age+{\alpha }_{4}Sale\_growth+{\alpha }_{5}Asset\\ & +{\alpha }_{6}Export+{\alpha }_{7}Internal\_finance+{\alpha }_{8}External\_debt+{\alpha }_{9}External\_equity+\\ & {\alpha }_{10}Collateral+{\alpha }_{11}Joint\_RD+{\alpha }_{12}Self\_RD+{\alpha }_{13}Market\_tech+{\alpha }_{14}Tech\_employee\\ & +{\alpha }_{15}Knowledge\_advice+{\alpha }_{16}Knowledge\_supplier+{\alpha }_{17}Knowledge\_Customer\\ & +{\alpha }_{18}Knowledge\_rival+{\alpha }_{19}Knowledge\_university+{\alpha }_{20}Knowledge\_association\\ & +{\alpha }_{21}Industry+{\alpha }_{22}Subsidy+\epsilon \end{array}$$

(4)

$$\begin{array}{ll}Subsidy\_y& ={\alpha }_0+{\alpha }_{1}State+{\alpha }_{2}Risk\_preference+{\alpha }_{3}Firm\_age+{\alpha }_{4}Sale\_growth+{\alpha }_{5}Asset\\ & +{\alpha }_{6}Export+{\alpha }_{7}Internal\_finance+{\alpha }_{8}External\_debt+{\alpha }_{9}External\_equity\\ & +{\alpha }_{10}Collateral+{\alpha }_{11}Reward+{\alpha }_{12}Industry+{\alpha }_{13}Innovation+\epsilon \end{array}$$

(5)

where Reward is a dummy variable with one if the firm has got a national reward from the government. We use Reward to control for the potential connection between firm production and government recognition, deriving from the question “whether the firm has received a national reward from the government for the firm’s product over the last three years?” [11].

As previously discussed, survey-based analyses may suffer from selection bias. Additionally, uncontrolled factors could violate the assumption that the variables of interest are independent from the error terms, even when control variables are included. To mitigate these concerns, many studies on the effectiveness of government subsidies employ propensity score matching (PSM) [1,25]. Therefore, to enhance the robustness of the results, we employed the PSM analysis method in the robustness checks section.

Our study employs three econometric approaches to address distinct challenges in analyzing the impact of government subsidies on SME innovation. The Heckman Two-Stage Model corrects sample selection bias in survey-based data through inverse Mills ratio estimation. Structural Equation Modeling (SEM) resolves endogeneity from bidirectional subsidy-innovation relationships by simultaneously estimating subsidy allocation and innovation. Propensity Score Matching (PSM) controls observable heterogeneity using 1:1 nearest-neighbor matching with post-matching diagnostics confirming eliminated group differences.

These methods form a “defense-in-depth” strategy: The Heckman model addresses unobservable selection factors, SEM clarifies causal pathways amid reciprocal relationships, while PSM validates robustness through quasi-experimental design. This triangulation is particularly crucial given SMEs’ heterogeneous responses to subsidies, as their complementary strengths collectively enhance validity—Heckman ensures baseline estimates account for hidden biases, SEM disentangles cherry-picking effects, and PSM provides counterfactual evidence.

4. Empirical Findings and Discussion

In this section, we present our empirical results and discuss their implications. We begin by analyzing the factors influencing the allocation (selection) of government subsidies and their relationship with firms’ financial constraints.

4.1. Baseline Results

We start by presenting the results of our model (Equation (1)), which estimates the probability of an SME receiving a government subsidy.

Table 2. Subsidy obtaining and financial constraints.

Dep V	Probit Model			Heckman Models
	Grant_d	Loan_d	Tax_d	Finance_Difficulty		Finance_Difficulty		Finance_Difficulty
	Coef.	Coef.	Coef.	Coef.	Odds Ratio	Coef.	Odds Ratio	Coef.	Odds Ratio
Grant_dummy				0.306 ** (0 153)	0.452
Loan_dummy								0.687 *** (0.198)	0.510
Tax_dummy						0.263 * (0.148)	0.640
state	0.176 (0.285)	0.254 (0.303)	0.731 *** (0.252)	−0.541 (0.360)		−0.606 (0.417)	−0.131	0.247 (0.370)
Risk-preference	0.135 (0.202)	0.192 (0.229)	0.250 (0.190)	−0.076 (0.247)		−0.085 (0.235)	−0.559	−0.321 (0.265)
Firm_age	−0.044 *** (0.011)	−0.035 *** (0.012)	−0.031 *** (0.010)	0.043 ** (0.022)	0.040	0.036 ** (0.016)	0.096	0.018 (0.016)
Sale_growth	0.129 ** (0.060)	0.157 ** (0.065)	0.078 (0.057)	0.100 (0.098)		0.125 * (0.075)		0.384 *** (0.093)	0.374
Asset	0.058 ** (0.028)	−0.012 (0.031)	−0.041 (0.026)	−0.056 (0.043)		−0.031 (0.032)	−0.267	−0.121 *** (0.032)	−0.100
Export	0.229 ** (0.111)	0.367 *** (0.124)	0.239 ** (0.106)	−0.191 (0.163)		−0.178 (0.153)	−0.393	−0.016 (0.177)
Internal_finance	−0.328 *** (0.110)	−0.238 * (0.123)	−0.228 ** (0.105)	0.135 (0.196)		0.097 (0.150)		−0.661 *** (0.155)	−0.598
External_debt	0.472 *** (0.129)	0.497 *** (0.150)	0.206 * (0.117)	0.303 (0.256)		0.373 ** (0.155)	−1.015	−0.129 (0.214)
External_equity	0.454 *** (0.141)	0.558 *** (0.146)	0.359 *** (0.136)	0.238 (0.264)		0.290 (0.218)		0.877 *** (0.254)	0.862
Collateral	−0.256 ** (0.111)	−0.334 *** (0.125)	−0.076 (0.106)	0.027 (0.171)		−0.018 (0.125)		−0.400 ** (0.168)	−0.461
Joint_RD	0.179 (0.128)	0.247 * (0.138)	0.145 (0.122)	0.391 ** (0.182)	0.384	0.401 ** (0.161)	0.389	0.791 *** (0.185)	0.808
Self_RD	−0.056 (0.125)	−0.261 * (0.144)	−0.047 (0.118)	0.331 ** (0.140)	0.408	0.332 ** (0.139)		0.210 (0.165)
Market_tech	0.026 (0.037)	0.011 (0.041)	0.044 (0.034)	0.055 (0.043)		0.044 (0.044)		0.016 (0.044)
Tech_employee	−0.053 (0.038)	−0.106 ** (0.042)	0.001 (0.037)	0.168 *** (0.050)	0.170	0.158 *** (0.044)	0.125	0.042 (0.056)
Knowledge_ supplier	0.074 (0.104)	0.071 (0.117)	0.067 (0.099)	0.305 ** (0.120)	0.340	0.308 ** (0.119)		0.002 (0.130)
Knowledge_ customer	−0.152 (0.108)	−0.126 (0.122)	−0.203 ** (0.102)	−0.143 (0.139)		−0.150 (0.138)		−0.171 (0.134)
Knowledge_ rival	−0.001 (0.123)	0.034 (0.138)	0.109 (0.115)	−0.077 (0.130)		−0.104 (0.136)		0.040 (0.149)
Knowledge_ university	0.037 (0.151)	0.316 * (0.163)	0.284 ** (0.140)	0.070 (0.180)		0.019 (0.208)		0.081 (0.220)
Knowledge_ association	0.538 *** (0.121)	0.320 ** (0.136)	0.368 *** (0.119)	−0.229 (0.289)		−0.127 (0.201)	−1.416	−0.290 (0.188)	−0.432
Knowledge_ advice	−0.007 (0.118)	−0.013 (0.134)	−0.128 (0.115)	0.059 (0.135)		0.082 (0.144)		0.070 (0.145)
Invmill				−0.278 (0.562)		−0.164 (0.460)		−0.205 (0.370)
reward	0.337 *** (0.127)	0.410 *** (0.119)	0224 *** (0.103)
cons	−1.787 *** (0.358)	−1.493 *** (0.390)	−1.257 *** (0.330)
Cut1				−2.317 * (1.328)	−2.414	−1.982 ** (0.921)	−6.440	−1.150 (0.861)	−1.647
Cut2				−0.818 (1.321)		−0.491 (0.912)	−3.625	1.621 * (0.862)
Cut3				0.025 (1.321)		0.339 (0.912)		3.192 *** (0.873)	2.709
Cut4				1.674 (1.322)	1.607	1.976 ** (0.914)		4.801 *** (0.912)	4.256
Industry FE	yes	yes	yes	yes	yes	yes	yes	yes	yes
obs	1095.000	1072.000	1103.000	1095.000	1072.000	1103.000	1095.000	1072.000	1103.000
LR chi2	186.630 ***	157.490 ***	119.370 ***	103.430 ***	107.050 ***	98.560 ***	267.200 ***	247.930 ***	252.520 ***
Pseudo R2	0.183	0.194	0.113	0.028	0.030	0.026	0.108	0.102	0.101

Notes: Standard errors in parentheses. Asterisks denote significance at the * 0.10, ** 0.05 and *** 0.01 level.

, columns 2, 3, and 4 show the probit model estimates for obtaining grants, loans, and tax credits, respectively.

The findings indicate that government ownership does not significantly impact the likelihood of receiving grants or loans. However, it has a positive and significant effect (at the 1% level) on tax credits. This suggests that while government-controlled SMEs are not more likely to receive direct financial support, they have a greater advantage in accessing indirect subsidies such as tax benefits.

Since 2007, when China adopted its “indigenous innovation” strategy, government subsidies—especially those initiated by the Ministry of Finance (MOF) and the Ministry of Science and Technology (MOST), such as SME technology innovation funds—have become increasingly transparent and specialized. Although external experts conduct evaluations, final grant decisions are made by program officials, with overarching control retained by higher-level authorities such as MOST. Table 2 shows that state-owned SMEs are 25.2% more likely to obtain tax credits than their non-state-owned counterparts. Additionally, the significant coefficient for the Reward variable suggests that government recognition of a firm’s products enhances its chances of securing subsidies.

Regarding firm characteristics, younger firms are more likely to receive government subsidies, aligning with the widespread incubator policies supporting high-tech SMEs in China [51]. Export-oriented SMEs are also significantly more likely to receive grants, loans, and tax credits, reflecting the emphasis on export-promoting policies in China. Additionally, larger firms are more likely to receive grants, while high-growth firms have better access to both grants and loans. As expected, SMEs relying primarily on internal financing are less likely to obtain government subsidies, highlighting the role of external financial constraints in subsidy allocation.

4.2. Subsidies and Financial Constraints

In our analysis, we use Finance difficulty as an ordinal dummy variable, derived from a seven-point Likert scale question in the survey (“How severe are your firm’s current financial constraints? 1 = Extremely easy to 7 = Extremely difficult”). As shown in Table 2, this ordinal variable evaluates policy effectiveness through three key aspects: credit availability, liquidity pressure, and cost burden. Compared to traditional financial indicators, this measure offers two important advantages: first, it helps identify institutional barriers that are not reflected in financial statements; second, it more accurately reflects actual corporate decision making processes. Table 2, columns 5–10 present the results of the ordinal logit models, where columns 5, 7, and 9 show the estimated coefficients for the effects of grants, loans, and tax credits on SME financing difficulties, respectively, while columns 6, 8, and 10 report the corresponding odds ratios.

Our findings indicate that all three forms of government subsidies—grants, loans, and tax credits—are positively associated with self-reported financial difficulties among SMEs. This suggests that, despite receiving financial support, SMEs still experience funding constraints, particularly in innovation-related activities. These positive effects provide further evidence that SMEs face significant financial barriers, especially when engaging in innovation. For financially constrained innovative SMEs, the additionality effect is more likely to occur than crowding out, as they rely on external funding due to limited internal resources.

Furthermore, we find that both joint and independent innovation activities, as well as hiring skilled employees, exacerbate financial difficulties for SMEs. This suggests that acquiring and integrating innovation knowledge from external sources, such as suppliers, may increase financial strain on firms. Lastly, firm age is positively correlated with financing difficulties, consistent with Czarnitzki & Delanote [52]. Older firms may face persistent challenges in accessing finance, potentially due to structural inefficiencies or a declining ability to attract investment.

4.3. Additionality or Crowding Out?

4.3.1. Subsidies and Innovation Inputs

Table 3. Subsidies and innovation inputs.

Dep V	RD_Intensity
	Coef.	Odds Ratio	Coef.	Odds Ratio	Coef.	Odds Ratio	Coef.	Odds Ratio	Coef.	Odds Ratio	Coef.	Odds Ratio
Grant_dummy	0.834 *** (0.160)	2.303					0.794 *** (0.161)	2.212
Loan_dummy			0.688 *** (0.182)	1.989					0.685 *** (0.182)	1.982
Tax_dummy					0.893 *** (0.154)	2.441					0.876 *** (0.154)	2.400
State*grant_ dummy							0.308 * (0.165)	1.360
State*loan_ dummy									0.116 (0.199)
State*tax_ dummy											0.404 ** (0.190)	1.497
state	−0.273 (−0.369)		−0.198 (−0.384)		−0.791 * (−0.432)	0.453	−0.736 (−0.449)		−0.275 (−0.406)		−1.315 *** (−0.497)	0.268 (−0.369)
Invmill	−1.773 *** (−0.557)	0.169	−1.016 *** (−0.339)		−1.209 *** (−0.456)	0.298	−1.731 *** (−0.557)	0.177	−1.033 *** (−0.341)	0.125	−1.284 *** (−0.458)	0.277 (−0.557)
Industry FE	Yes	yes	yes	yes	yes	yes	yes	yes	yes	yes	yes	yes
Control variables
obs	1095.000	1095.000	1072.000	1072.000	1103.000	1103.000	1095.000	1095.000	1072.000	1072.000	1103.000	1103.000
LR chi2	391.860 ***	391.860 ***	372.760 ***	372.760 ***	372.760 ***	399.980 ***	399.980 ***	395.320 ***	395.320 ***	373.100 ***	373.100 ***	404.630 ***
Pseudo R2	0.107	0.107	0.104	0.104	0.108	0.108	0.108	0.108	0.108	0.104	0.104	0.1101

Notes: Standard errors in parentheses. Asterisks denote significance at the * 0.10, ** 0.05 and *** 0.01 level.

presents the ordinal logit model estimation results for the effects of government subsidies on innovation inputs. The findings indicate that subsidies significantly and positively influence R&D intensity. Among the three types of government support, tax credits have the strongest effect, with an odds ratio of 2.441, meaning that SMEs receiving tax credits are 144.1% more likely to increase innovation expenditures. In comparison, grants (odds ratio = 2.303) and loans (odds ratio = 1.989) exhibit slightly weaker effects.

4.3.2. Subsidies and Innovation Outputs

Table 4. Subsidies and innovation outputs.

Dep V	Innovative_Sales
	Coef.	EXP Marginal Eff-1	Coef.	EXP Marginal Eff-1	Coef.	EXP Marginal Eff-1	Coef.	EXP Marginal Eff-1	Coef.	EXP Marginal Eff-1	Coef.	EXP Marginal Eff-1
Grant_dummy	1.819 *** (0.168)	5.165					1.796 *** (0.170)	5.025
Loan_dummy			1.261 *** (0.201)	2.528					1.269 *** (0.201)	2.557
Tax_dummy					1.065 *** (0.165)	1.900					1.071 *** (0.165)	1.918
State*grant_ dummy							0.142 (0.173)
State*loan_ dummy									−0.187 (0.228)
State*tax_ dummy											−0.091 (0.202)
state	−0.323 (0.375)		−0.209 (0.396)		−0.364 (0.462)		−0.502 (0.434)		−0.139 (0.405)		−0.273 (0.504)
Invmill	−1.130 * (0.594)	−0.676	−0.643 * (0.381)	3.375	−0.277 (0.510)		−1.110 * (0.595)	−0.670	−0.617 (0.382)		−0.263 (0.512)
cons	2.756 ** (1.401)	14.736	1.476 * (0.889)		0.554 (1.003)		2.721 * (1.402)	14.195	1.414 (0.892)		0.518 (1.007)
Industry FE	yes		yes		yes		yes		yes		yes
Control variables
obs	1095.000		1072.000		1103.000		1095.000		1072.000		1103.000
F test	13.370 ***	640,496.200	10.620 ***	40,944.610	10.480 ***	35,595.410	13.050 ***	465,095.400	10.360 ***	31,570.180	10.230 ***	27,721.510
R2	0.330		0.280		0.283		0.331		0.281		0.283

Notes: Standard errors in parentheses. Asterisks denote significance at the * 0.10, ** 0.05 and *** 0.01 level.

reports the results on the impact of government subsidies on innovation outputs. We find that all three subsidy types positively influence innovation, with grants showing the largest impact. Accordingly, we fail to reject Hypothesis 1, which posits that government subsidies promote innovation among high-tech SMEs rather than crowd out private investment.

The estimated marginal effects show that SMEs receiving grants achieve 4.165 times higher innovation sales than those without. Loans have a smaller effect (2.528), while tax credits have the weakest (1.900). These differences suggest that direct subsidies (e.g., grants and loans) yield quicker and stronger effects on innovation output compared to indirect subsidies (e.g., tax credits).

Regarding tax credits, we observe a paradoxical “high-input but low-output” effect, which can be attributed to several factors. First, tax credits directly reduce corporate tax burdens, significantly increasing disposable funds available for R&D investment, accounting for their strong input effects. However, the weaker output effects may stem from: (1) Time lag effects: Tax credits typically rebate historical R&D expenditures, while commercializing innovation outcomes requires a 2–3 year cycle; (2) Regulatory leniency: Compared to grants that require innovation performance evaluation, tax credits have lower eligibility thresholds, potentially encouraging “strategic R&D” rather than substantive innovation; (3) Policy transmission differences: Direct subsidies (e.g., grants) can precisely support specific projects, whereas tax credits, as universal policies, provide limited marginal incentives for innovation outputs. These findings suggest that policy design should balance short-term input stimulation with long-term output assessment.

4.3.3. State Ownership, Subsidies, and Innovation

Beyond increasing the likelihood of obtaining government subsidies, state ownership also has a positive moderating effect on the relationship between subsidies and innovation inputs, particularly for grants and tax credits (Table 3, columns 8 and 12). State-owned SMEs show higher R&D intensity, with odds ratios of 1.36 for grants and 1.497 for tax credits. This supports Hypothesis 2a, which suggests that state ownership positively moderates the effect of subsidies on R&D spending. However, state-owned SMEs without access to grants or tax credits appear more conservative in R&D investment. Table 4, however, indicates that state ownership does not significantly influence innovation outputs, leading us to reject Hypothesis 2b. This finding implies that state-owned SMEs struggle to efficiently convert innovation inputs into outputs, possibly due to soft-budget constraints and inefficiencies commonly associated with state ownership.

4.3.4. Control Variables

Firm Age & Growth: Mature firms are more likely to invest in R&D (odds ratio ≈ 1.04), though older SMEs do not necessarily achieve higher innovation sales. Growing firms invest more in R&D, but sales growth only significantly correlates with innovation output in the tax credit model.

Firm Size: While firm size does not show a consistent relationship with R&D intensity, it does positively correlate with innovation sales, suggesting that larger firms are more successful in translating R&D investments into marketable innovations.

Export Orientation: Exporting SMEs exhibit a more conservative attitude toward R&D investment.

Collateral: Firms with greater access to collateral tend to invest more in R&D, likely because collateral helps mitigate adverse selection and moral hazard issues in financing [19].

Independent vs. Joint R&D: Independent R&D drives higher demand for R&D spending and stronger sales growth than joint R&D.

Knowledge Sources: Knowledge acquired from industry associations consistently enhances innovation sales across all models.

4.4. Endogeneity and Cherry Picking

To address potential endogeneity in the allocation of government subsidies—specifically, the cherry-picking effect—we employ a structural equation model (SEM) following Equations (4) and (5). The three-stage SEM estimation results are presented in

Table 5. Endogeneity and cherry picking.

Dep V	Model A		Model B		Model C		Model D		Model E		Model F
	RD_Intensity	Grants	RD_Intensity	Loans	RD_Intensity	Tax Credits	Innovative_Sales	Grants	Innovative_Sales	Loans	Innovative_Sales	Tax Credits
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)
Grants	1.040 *** (0.365)						1.228 *** (0.355)
Loans			1.525 *** (0.533)						1.613 *** (0.560)
Tax credits					2.639 * (1.482)						2.802 * (1.470)
RD_intensity		0.153 ** (0.073)		0.116 * (0.066)		0.124 * (0.068)
Innovative_ sales								0.414 *** (0.083)		0.235 *** (0.075)		0.244 *** (0.076)
state	−0.595 (0.413)	0.281 (0.259)	−0.274 (0.435)	0.042 (0.243)	−2.013 * (1.163)	0.699 *** (0.250)	−0.341 (0.401)	0.239 (0.247)	0.020 (0.459)	−0.011 (0.239)	−1.796 (1.155)	0.641 *** (0.243)
Reward		0.466 *** (0.131)		0.234 (0.097)		0.110 (0.100)		0.239 (0.131)		0.099 * (0.097)		0.050 (0.098)
cons	0.542 (0.420)	−0.459 ** (0.231)	0.324 (0.428)	−0.255 (0.216)	1.028 (0.935)	−0.402 * (0.224)	1.055 *** (0.395)	−0.684 *** (0.227)	0.737 * (0.444)	−0.357 * (0.215)	0.123 (0.925)	−0.508 ** (0.219)
Industry FE	yes	yes	yes	yes	yes	yes	yes	yes	yes	yes	yes	yes
Control variables
obs	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000	1103.000
LR chi2	262.580 ***	234.680 ***	228.920 ***	166.500 ***	289.310 ***	133.800 ***	420.560 ***	274.210 ***	313.190 ***	176.890 ***	506.110 ***	147.850 ***
R2	−0.187	0.186	−0.736	0.136	−3.099	0.119	0.179	0.249	−0.236	0.155	−1.682	−1.682

Notes: Standard errors in parentheses. Asterisks denote significance at the * 0.10, ** 0.05 and *** 0.01 level.

. Models A, B, and C (columns 2–6) examine whether high R&D intensity attracts government subsidies, i.e., whether policymakers selectively allocate funding to firms already investing heavily in innovation. Models D, E, and F (columns 7–12) assess whether innovation sales influence subsidy allocation.

Table 5 reveals a two-way relationship between R&D intensity and government subsidies. All three subsidy types (grants, loans, and tax credits) positively influence R&D intensity, while firms with higher R&D intensity are more likely to receive subsidies. Notably, after accounting for endogeneity, the positive effect of subsidies on R&D intensity becomes stronger. This suggests that additionality plays a more dominant role in innovation inputs than cherry-picking. Interestingly, tax credits exhibit the largest increase in coefficient values, though their p-value rises to 0.1, indicating a weaker statistical significance. These findings support Hypothesis 3a.

Models D, E, and F (columns 7–12 in Table 5) confirm a two-way effect between innovation output and government subsidies. Firms with higher innovation sales are more likely to receive subsidies, reinforcing Hypothesis 3b. However, unlike the results for innovation inputs, the coefficient values of grants decrease in the SEM model compared to the Heckman models (see columns 2 and 4 in Table 4). Conversely, tax credits show a significant increase in coefficients in the SEM model (see column 6 in Table 4 and column 12 in Table 5). We argue that grants exhibit more cherry-picking characteristics in innovation outputs than in innovation inputs, meaning that grants are more likely to be awarded to SMEs that have already demonstrated innovation success. In contrast, loans show a stronger additionality effect than a cherry-picking effect. While tax credits exhibit a coefficient increase in additionality effects, their p-value increases from 0.01 to 0.1, suggesting reduced statistical confidence. By contrast, the cherry-picking coefficient maintains a p-value of 0.01 (see columns 11 and 12 in Table 5). Given these results—along with the decreasing coefficient for grants—we accept Hypothesis 3c, except for the case of loans.

Comparing Table 5 with earlier results (Table 2, Table 3 and Table 4), we observe consistent signs and significance for most control variables. However, some variables show notable differences:

Firm age, external equity, and external debt become insignificant in the SEM model. Firm growth exhibits a negative effect on receiving grants, suggesting that fast-growing SMEs are less likely to obtain grants once R&D intensity is controlled for. These findings highlight the complex nature of subsidy allocation and suggest that some funding mechanisms favor established, innovation-intensive firms over rapidly growing SMEs.

4.5. Robustness Test

To enhance the robustness of our results, we implement PSM analysis, treating government subsidy as a treatment variable. Treatment group: Firms that received government subsidies (grants, loans, or tax credits). Control group: Firms that did not receive any government subsidies. Using propensity scores obtained from Probit models (see Table 2), we perform nearest-neighbor 1:1 matching, where each treated firm is paired with the most similar untreated firm based on observable characteristics. This method ensures that the control and treatment groups are comparable, reducing heterogeneity bias in our regression estimates.

To validate the assumptions of common support and balancing property in PSM, we examine the mean differences between the treatment and control groups before and after matching.

Table 6. A comparison between matched and unmatched samples.

Variable	Unmatched Matched	Mean		% Change		t-Test
		Treated	Control	% Bias	|bias|	t	p > |t|
state	U	0.053	0.021	17.200		2.800	0.005
	M	0.053	0.565	−1.900	89.100	−0.180	0.854
Risk-preference	U	0.074	0.052	8.900		1.350	0.177
	M	0.074	0.060	5.800	35.100	0.670	0.503
Firm_age	U	2.904	4.379	−27.600		−3.620	0.000
	M	2.905	3.313	−7.600	72.300	−1.070	0.285
Sale_growth	U	2.131	1.910	25.400		3.820	0.000
	M	2.131	2.187	−6.500	74.400	−0.730	0.463
Asset	U	6.983	6.421	26.000		3.730	0.000
	M	6.983	6.971	0.600	97.800	0.070	0.945
Export	U	0.406	0.271	28.900		4.300	0.000
	M	0.406	0.378	6.000	79.200	0.690	0.492
Internal_finance	U	0.618	0.704	−18.100		−2.660	0.008
	M	0.618	0.594	5.200	71.000	0.600	0.548
External_debt	U	0.788	0.704	19.400		2.750	0.006
	M	0.788	0.795	−1.600	91.600	−0.210	0.836
External_equity	U	0.201	0.089	32.300		5.110	0.000
	M	0.201	0.194	2.000	93.700	0.210	0.833
Collateral	U	0.548	0.560	−2.400		−0.350	0.725
	M	0.548	0.579	−6.400	−163.900	−0.760	0.446
Joint_RD	U	0.237	0.168	17.100		2.560	0.011
	M	0.237	0.208	7.100	58.700	0.810	0.420
Self_RD	U	0.310	0.223	19.900		2.970	0.003
	M	0.310	0.300	2.400	87.900	0.270	0.785
Market_tech	U	4.625	4.423	14.000		2.030	0.043
	M	4.625	4.473	10.500	24.900	1.230	0.219
Tech_employee	U	4.891	4.932	−3.100		−0.460	0.646
	M	4.891	4.813	5.800	−88.500	0.670	0.502
Knowledge_supplier	U	0.565	0.548	3.600		0.520	0.604
	M	0.565	0.583	−3.600	0.800	−0.420	0.671
Knowledge_customer	U	0.534	0.545	−2.300		−0.340	0.737
	M	0.534	0.519	2.800	−22.300	0.340	0.737
Knowledge_rival	U	0.247	0.259	−2.600		−0.370	0.710
	M	0.247	0.194	12.200	−373.800	1.520	0.129
Knowledge_university	U	0.159	0.107	15.200		2.310	0.021
	M	0.159	0.191	−9.400	38.500	−0.990	0.320
Knowledge_association	U	0.336	0.155	42.900		6.670	0.000
	M	0.336	0.360	−5.900	86.300	−0.620	0.538
Knowledge_advice	U	0.279	0.232	10.900		1.600	0.109
	M	0.279	0.276	0.800	92.600	0.090	0.925

reports these results, where U (unmatched) and M (matched) denote samples before and after PSM, respectively. Before matching, firms in the treatment group (subsidy recipients) significantly differ from those in the control group in terms of ownership structure, firm age, sales growth, assets, export status, and other characteristics. These differences indicate that subsidy recipients are not directly comparable to non-recipients, which could introduce selection bias in regression models. After matching, however, the differences disappear, confirming that the two groups are now statistically comparable.

In

Table 7. ATT estimates (differences in the means of matched and unmatched samples).

Innovation	Treatment	Sample	Treated	Controls	Difference	S.E.	T-Stat	Num
RD_intensity	Subsidy	Unmatched	2.752	2.132	0.620	0.072	8.530	1103.000
		ATT	2.752	2.300	0.452	0.103	4.360	283.000
	Grant	Unmatched	2.860	2.171	0.689	0.083	8.230	1103.000
		ATT	2.860	2.528	0.332	0.143	2.300	193.000
	Loan	Unmatched	2.844	2.214	0.630	0.098	6.410	1103.000
		ATT	2.844	2.614	0.230	0.162	1.410	135.000
	Tax	Unmatched	2.832	2.170	0.662	0.082	8.060	1103.000
		ATT	2.832	2.305	0.527	0.123	4.260	203.000
Innovative_sales	Subsidy	Unmatched	4.980	3.028	1.952	0.148	13.160	1103.000
		ATT	4.980	3.398	1.582	0.235	6.710	283.000
	Grant	Unmatched	5.560	3.098	2.462	0.167	14.690	1103.000
		ATT	5.560	4.045	1.515	0.317	4.780	193.000
	Loan	Unmatched	5.063	3.314	1.749	0.205	8.490	1103.000
		ATT	5.063	3.887	1.176	0.323	3.640	135.000
	Tax	Unmatched	4.653	3.275	1.378	0.174	7.880	1103.000
		ATT	4.653	3.718	0.935	0.276	3.380	203.000

, we compare innovation inputs and outputs between: Treated firms (firms that received subsidies). Matched firms (firms in the control group with similar characteristics). Our findings reveal that: Firms receiving grants, loans, or tax credits exhibit higher levels of innovation inputs and outputs compared to their matched counterparts. Although the magnitude of differences is smaller than in the unmatched sample, the results still indicate a positive additionality effect of government subsidies. This confirms that government support effectively stimulates innovation activities, reinforcing our main findings. Overall, the PSM results provide strong evidence that government subsidies—including grants, loans, and tax credits—enhance both innovation inputs and outputs, validating our conclusions about the role of public funding in fostering firm innovation.

4.6. Further Discussion

Our empirical findings reveal a complex relationship between government subsidies and innovation in high-tech SMEs. The study contributes to the literature in the following aspects:

4.6.1. Additionality vs. Crowding-Out Effect

The positive correlation between government subsidies, innovation inputs, and innovation outputs supports the additionality theory [7], particularly emphasizing the role of loan-based subsidies. This contrasts with the crowding-out hypothesis demonstrated by Dimos and Pugh [17], indicating that in the context of Chinese high-tech SMEs, government subsidies primarily serve as an incentive for corporate innovation rather than displacing private investment.

4.6.2. The State Ownership Paradox

In terms of innovation inputs, state-owned enterprises (SOEs) exhibit higher additionality effects. However, given their exposure to soft budget constraints, their innovation output efficiency tends to be relatively low, aligning with the findings of Kornai [42]. This suggests that while SOEs receive substantial subsidies and invest more in R&D, their innovation outcomes may not be as efficient as those of private firms.

4.6.3. The “Cherry-Picking” Mechanism

The two-way relationship between government subsidies and corporate innovation confirms the “cherry-picking” behavior described by Czarnitzki and Licht [10]. Additionally, the presence of a “reward effect” in grant-based subsidies aligns with public choice theory [17], where policymakers may selectively favor firms perceived as more innovative. Meanwhile, the ex-post nature of tax credits may lead firms into a cycle of marginal improvements, further influenced by government selection biases.

4.6.4. Boundary Conditions

The additionality effect of government subsidies is closely linked to China’s unique innovation policies. However, in emerging economies, where legal and regulatory frameworks remain underdeveloped, government subsidies may be more susceptible to cherry-picking behavior, potentially leading to suboptimal allocation of public resources.

This study theoretically advances the understanding of subsidy-innovation dynamics by empirically validating the dominance of additionality effects in China’s unique institutional context, while revealing three critical paradoxes: (1) loan-based subsidies’ efficacy versus crowding-out absence, (2) SOEs’ input-output efficiency disconnect under soft budget constraints, and (3) policy-driven cherry-picking that distorts both firm behavior and subsidy allocation mechanisms.

5. Concluding Remarks

Our analysis of Chinese high-tech SMEs reveals three core insights about government subsidies’ innovation effects:

5.1. Net Positive Additionality

Subsidies significantly boost both innovation inputs (R&D intensity) and outputs (new product sales), with loans demonstrating the strongest and most consistent additionality effects across metrics. Grants, while effective, show susceptibility to cherry-picking—coefficients for output effects drop by 32.5% (from 1.819 to 1.228) when accounting for selection bias.

5.2. Financial Constraints Persist

Despite promoting innovation, subsidies fail to alleviate systemic financial pressures on SMEs. They primarily act as seed funding for early-stage projects rather than solving structural financing gaps.

5.3. Ownership Dynamics

State ownership enhances input additionality but correlates with lower innovation efficiency, as SOEs prioritize compliance over market-driven R&D.

To address these challenges, optimize subsidy allocation, and mitigate cherry-picking behavior, policymakers should consider the following measures:

5.4. Applying Differentiated Subsidy Instruments

To optimize innovation incentives through differentiated subsidy instruments. Prioritize loan-based subsidies, which demonstrate significantly stronger input and output additionality effects compared to grants and tax credits. Implement an “innovation milestone” evaluation mechanism for grants, requiring predefined thresholds for patent counts or R&D personnel ratios to mitigate cherry-picking in subsidy allocation. The proposed measures include implementing a three-tiered interest subsidy policy (provincial/national/key strategic projects receiving 40%/60%/80% subsidies respectively), requiring 1:1 matching funds from enterprises for subsidies exceeding 5 million RMB, and disbursing 30% of grants in installments contingent upon meeting either of the following conditions: ≥15% annual growth in R&D personnel ratio or ≥3 patent applications per 1 million RMB in funding.

5.5. Establishing Subsidy-Finance Linkage Mechanisms

Address financing constraints for high-tech SMEs by creating subsidy-finance linkage mechanisms. Link subsidy eligibility to corporate credit ratings, granting prioritized credit access and preferential interest rates to firms listed on the “Technology Credit Whitelist.” Concurrently, local governments should establish matched risk-sharing funds to incentivize commercial banks to develop financing products tailored for SME innovation. The proposed measures include offering interest rate discounts of up to 2% for enterprises maintaining ≥20% annual R&D growth for three consecutive years, providing banks with 40% first-loss guarantee coverage, and establishing blockchain-based automated monitoring platforms for subsidy distribution.

5.6. Implementing Ownership-Specific Efficiency Evaluations

Enhance overall innovation efficiency through ownership-differentiated evaluation systems. For state-owned enterprises (SOEs), enforce mandatory “input-output conversion rate” benchmarks pegged to industry averages. For private firms, adopt a “breakthrough exemption” mechanism—waiving subsequent evaluations upon achieving milestone technological breakthroughs—to preserve market-driven R&D flexibility. The proposed measures include mandating SOEs to conduct compulsory benchmarking against the top three private enterprises in their sector, while allowing private firms to obtain a two-year evaluation exemption for breakthrough technologies and granting a 50% tax credit bonus for PCT patent applications.

While this study employs three econometric methods to mitigate biases, the following limitations warrant attention:

5.7. Self-Reporting Bias

R&D expenditures and sales of new products rely on self-reported data, which may involve systematic overreporting by firms.

5.8. Latent Variable Confounding

All variables were collected through a single survey instrument, leaving potential confounding effects from unobserved latent variables in analyzing the subsidy-innovation relationship.

5.9. External Validity Constraints

As the sample focuses exclusively on high-tech industries, findings may not generalize to traditional manufacturing sectors.

While this study is based on 2015 data with certain temporal limitations, recent related research demonstrates that our findings remain valid. The core conclusions not only enrich theoretical understanding of government subsidies and corporate innovation, but also maintain relevance for current policy optimization. Future research will incorporate more recent data for comparative analysis. We will also expand the mechanism analysis to examine specific pathways through which subsidies affect innovation, particularly from the perspective of alleviating financial constraints. Furthermore, we plan to investigate industry heterogeneity to reveal differences in subsidy effects across sectors. Finally, when data permits, we will enhance the historical comparative value of this study by incorporating international comparisons with other countries/regions, thereby highlighting the distinctive features of China’s policies and their broader implications.