Replication Study of “The Use of Variance Decomposition in the Investigation of CEO Effects Using Machine Learning: How Large Must the CEO Effect Be to Rule Out Chance?”

The extent to which CEOs impact corporate performance remains one of the most contentious issues in strategic management. Previous research, most notably Fitza (2014), used ANOVA to decompose variance and showed that CEO effects are minimal after accounting for randomness. However, linear models ignore complex relationships and delayed leadership influence. This work replicates Fitza’s analysis and advances the argument by using XGBoost, a non-linear machine learning technique, on an enhanced dataset spanning 1999–2024 that includes Compustat, ExecuComp, and BoardEx data. Our model integrates comprehensive CEO traits (e.g., tenure, education, remuneration) as well as firm-level financial indicators, resulting in lagged variables across an eight-year period that capture both immediate and cumulative effects.The XGBoost model shows strong predictive performance (train R² = 0.675; test R² = 0.546), indicating significant explanatory power for firm performance. According to feature importance analysis, firm-level financial indicators dominate short-term prediction over CEO-related variables. A combined lag-year and entity-type study reveals that firm effects peak in the current year and then rapidly fall after two years, whereas CEO influence steadily grows until year three and then stabilizes over time. We can conclude from the data that business fundamentals are more important in predicting short-term outcomes, but the effect imparted by leadership requires more time to become apparent.This study provides a refined view of CEO influence by questioning past assumptions and showcases the potential of predictive modeling in corporate governance research.