Reproduce “Estimation in high-dimensional linear regression: Post-Double-Autometrics as an alternative to Post-Double-Lasso” on OpenResearch

About this paper

Sullivan Hué, Sébastien Laurent, Ulrich Aiounou, Emmanuel Flachaire

A new variable selection method, Post-Double-Autometrics, is introduced to enhance estimation accuracy in high-dimensional linear regression by integrating statistically-driven Autometrics into a double-selection framework. Monte Carlo simulations demonstrate its robustness and lower bias compared to Lasso-based methods, influencing findings in an economic growth application.

Key insights

• Autometrics' reliance on statistical significance for variable selection, rather than coefficient magnitude, ensures that relevant variables are retained even if their effects are small, thereby reducing OVB.
• Post-Double-Autometrics exhibits greater robustness to its tuning parameter (α) and covariate correlation structures compared to Lasso-based methods, which are highly sensitive to λ.
• The double-selection principle, when combined with statistically rigorous inference through Autometrics, yields more accurate and reliable causal effect estimates in complex data environments.

Summary by alphaXiv.