A0424
Title: Supervised dimension reduction for instrumental variables estimation with some invalid instruments
Authors: Kei Tsubotani - Graduate School of Doshisha University (Japan) [presenting]
Jun Tsuchida - Kyoto Womens University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: The instrumental variables (IVs) method is used to estimate treatment effects without bias when an unobserved confounder exists. This method relies on valid IVs that satisfy three assumptions: relevance, exclusion restriction, and independence. However, identifying valid IVs can be challenging without sufficient knowledge of the domain. To select valid IVs when the candidate IVs include invalid IVs that do not meet the exclusion restriction (i.e., variables that directly affect the outcome), several methods have been proposed by applying variable selection techniques commonly used in regression analysis. Meanwhile, supervised dimension reduction can be used to extract valid IVs because the subspace estimation method includes variable selection techniques. A method for estimating treatment effects using supervised dimension reduction is proposed under an assumed situation. Numerical experiments are conducted to evaluate the performance of the IV method using supervised dimension reduction methods from the viewpoint of the number of valid IVs and the strength of the relationship between the outcome and invalid IVs. The results reveal that the performance of the supervised dimension reduction methods was superior to that of the variable selection methods when the number of valid IVs was small.
