Phylogenetic comparative methods are essential tools for analyzing trait correlations among species while controlling for evolutionary dependencies. Phylogenetic eigenvector regression (PVR) uses eigenvectors (EVs) derived from the phylogenetic tree as covariates to remove shared evolutionary history. Although conceptually simple and independent of predefined evolutionary models, PVR has traditionally been viewed as inferior to phylogenetic generalized least squares (PGLS), which explicitly incorporates evolutionary models. Here, we systematically optimized the PVR framework by integrating robust statistical estimators (L1, L2, M, and MM) and developing a theoretically justified approach to selecting EVs, particularly by using the union of EVs from both traits rather than a single trait alone. Using extensive simulations, we compared optimized PVR methods to established approaches, including PGLS and phylogenetic independent contrasts (PIC), across a range of evolutionary scenarios, spanning strong, weak, and asymmetric phylogenetic signals as well as non-stationary shifts; and across multiple tree structures (balanced and ladder-shaped), branch-length scalings ({rho} = 0.1, 1, 2), and sample sizes (128 and 16 species). Our results challenge the prevailing assumption that PGLS universally outperforms PVR. While PGLS and standard PIC generally excelled when dependent traits exhibited strong phylogenetic signals, optimized PVR methods--particularly PVR-MM--demonstrated superior robustness under scenarios involving mixed or weak phylogenetic signals, as well as abrupt, non-stationary evolutionary shifts. Even the simplest optimized PVR variant (PVR-L2) exhibited inherent robustness, highlighting a fundamental advantage of phylogenetic eigenvectors in controlling evolutionary structure. Furthermore, PVR provides greater methodological flexibility compared to model-based methods like PGLS, as it is not constrained by predefined evolutionary models and can readily extend to advanced machine-learning approaches. Lastly, unlike PGLS, PVR effectively addresses the directional asymmetry inherited from ordinary least squares regression, allowing researchers to select predictors based on biological reasoning or measurement accuracy without compromising model validity. Overall, this study highlights that optimized PVR is not merely a competitive alternative to traditional methods, but also a flexible and robust framework that adapts well to a wide range of evolutionary scenarios and analytical needs.