The digital economy necessitates data-driven management and decisions. Machine learning (ML), a component of artificial intelligence, offers a powerful complement to traditional econometrics for strategic decision-making and pattern discovery. While econometrics provides statistically consistent and interpretable coefficients for hypothesis testing, ML's flexible functional forms reveal complex patterns in data, leading to new theories or refining existing ones. This research uses both approaches to examine latecomer firms' path-creating strategies and their impact on catch-up performance. Latecomers, disadvantaged in technology and markets, must overcome barriers set by first-movers. They can leverage incumbents' disadvantages or exploit unique latecomer advantages to catch up. The sectoral innovation system (SIS) framework (Malerba, 2002) is particularly relevant, encompassing interactions between knowledge base, technologies, markets, and actors. Lee and Lim (2001) identified three catch-up strategies: path-following, path-skipping, and path-creating. While research focuses on alternative paths, the conditions for successful path-creation remain unclear. Traditional case studies and linear regression provide limited qualitative and heterogeneous insights. This study addresses these gaps by combining econometric hypothesis testing with ML to uncover complex relationships and provide strategic decision-making guidance.

The literature highlights the challenges faced by latecomer firms in catching up with industry leaders. Incumbent firms often maintain path dependence, investing in existing technologies and neglecting new ones, thus creating opportunities for latecomers. However, successful catch-up often requires a strategic shift from path-following/skipping to path-creation, as evidenced by case studies of firms like China Telecom Systems Corporation. Existing research, however, lacks quantitative analyses to determine the factors affecting path-creation's success and to offer tailored guidance for firms. This study draws upon the sectoral innovation system (SIS) framework to identify key factors influencing the technological catch-up of latecomers, including technological regime features (appropriability, cumulativeness, opportunity, uncertainty, technological lifecycle time, degree of industrial innovation), market regime features (demand growth, demand fluctuation, industry concentration), and firm characteristics (age, size, R&D intensity, absorptive capacity, technology dependence, diversity, originality). This study addresses the limitations of previous research by combining traditional econometrics and machine learning to provide both average and heterogeneous insights, allowing for a more nuanced understanding of the complex relationships between path-creation, technological capability and catch-up performance.

This study used unbalanced panel data from 283 high-tech manufacturing firms listed on the Shanghai and Shenzhen stock exchanges from 2007 to 2019 (1805 observations). Data were obtained from CSMAR, Wind, and Incopat databases. Data cleaning involved eliminating ST-treated and delisted firms, addressing missing values (imputation or deletion), and removing outliers. The dependent variable, technological catch-up performance (TechCatchup), was measured as the ratio of a firm's patent applications to total industry patent applications in a given technology sector. Independent variables were constructed based on the SIS framework, including six technology regime factors, three market regime factors, and several firm-level characteristics (e.g., R&D intensity, absorptive capacity, path-creating strategy). Path-creating was measured using the self-citation rate of patents. Feature selection involved three algorithms: SelectKBest (F-statistic), Permutation Importance, and Random Forest. The top 18 features from each algorithm were identified, and their intersection (15 features) was used in subsequent analyses. OLS regression (ProcessMacro in Python) was used to test hypotheses regarding the direct and indirect effects of path-creating on TechCatchup, mediating role of technological capability and moderating effects of appropriability and cumulativeness. Causal machine learning (CausalML and scikit-learn in Python) was used to estimate the average treatment effect (ATE) and individual treatment effects (ITE) of the path-creating strategy, employing meta-learners (LGBMRegressor) for ITE analysis. Shapley values were calculated to assess feature importance for ITE, and a decision tree was constructed to visualize decision rules for effective path-creation.

Feature selection identified cumulativeness, technological capability, path-creating, and appropriability as key features. OLS regression showed a significant positive direct effect of path-creating on TechCatchup (b = 0.20, p < 0.05). Technological capability mediated this relationship, with the indirect effect being significant only at high levels of appropriability. Appropriability positively moderated the path-creating-capability relationship, supporting the hypothesis that protecting innovation enhances capacity building. Cumulativeness negatively moderated the direct effect of path-creating on TechCatchup (b = -0.02, p < 0.05), but positively moderated the capability-catch-up relationship, indicating that high cumulative knowledge aids firms with existing capabilities. Causal machine learning confirmed the positive ATE of path-creating on TechCatchup (ATE = 0.25, p < 0.05). ITE analysis revealed heterogeneity, with some firms experiencing negative effects. Shapley value analysis highlighted technological capability as a key driver of positive treatment effects, with appropriability and cumulativeness exhibiting complex interactions. The decision tree visualized 14 decision paths, illustrating that path-creating strategies are more likely to succeed for firms with high technological capabilities, but other factors such as appropriability and absorptive capacity play significant roles. The model also reveals circumstances under which path-creation might be successful even with low initial technological capability.

The findings confirm the positive relationship between path-creating strategies and technological catch-up, but also highlight the heterogeneity of this relationship. While the overall effect of path creation is positive, its effectiveness depends on a complex interplay of factors. High technological capability is a key enabler of success, while the impact of appropriability and cumulativeness is context-dependent. Appropriability, while generally beneficial, does not guarantee success. The decision tree underscores the complexities of successful path-creation, suggesting that firms need to carefully assess the technological environment (appropriability, cumulativeness, uncertainty), their internal capabilities (technological capability, absorptive capacity), and market conditions before adopting a path-creating strategy. The results contribute to the understanding of latecomer catch-up strategies by providing both aggregate and individual-level insights, offering actionable guidance for firms seeking to achieve technological leadership.

This study integrates econometrics and causal machine learning to analyze the mechanism and decision-making behind latecomers' path-creating strategies. It confirms the positive impact of path-creation on catch-up performance, mediated by technological capability, and moderated by appropriability and cumulativeness. However, the machine learning analysis reveals the heterogeneity of this relationship. Future research could explore other catch-up strategies, expand the scope of industries and countries, and investigate the dynamic interplay of factors influencing successful path creation over time.

The study's limitations include the use of patent data as a proxy for technological catch-up, which may not fully capture the complexity of technological advancements. The sample focuses on high-tech manufacturing firms listed in China, potentially limiting the generalizability of the findings to other sectors and national contexts. Future research could employ more comprehensive measures of technological capability and consider a broader range of industries and countries to enhance the study's generalizability.