On Statistical Inference in Factor Analysis: Polynomial-Time Verification of the Anderson-Rubin Condition

Beschreibung

In sparse factor models, identifiability of the variance decomposition has received little attention perhaps due to its difficulty. Anderson and Rubin (1956) famously establish identifiability under a rank assumption on a large number of submatrices of the factor loading matrix. This number is potentially exponential in the input dimensions and identifiability is therefore infeasible to verify for computers. In our paper, the computational complexity is reduced to the speedy inspection of just one special rotation, the generalized lower triangular (GLT) form. As part of exploratory factor analysis, our method is deployed to avoid nonsensical models independently of observation ordering in both Bayesian and frequentist contexts. Furthermore, a fully Bayesian sampling procedure is developed, which leverages on the GLT rotation while estimating the unknown number of latent factors. The procedure is applied to financial and economic data. Joint work with Sylvia Frühwirth-Schnatter.

Zeitraum	26 Juni 2022 → 1 Juli 2022
Ereignistitel	ISBA World Meeting
Veranstaltungstyp	Keine Angaben
Bekanntheitsgrad	International