The Annals of Statistics
- Ann. Statist.
- Volume 47, Number 2 (2019), 663-690.
Computation of maximum likelihood estimates in cyclic structural equation models
Software for computation of maximum likelihood estimates in linear structural equation models typically employs general techniques from nonlinear optimization, such as quasi-Newton methods. In practice, careful tuning of initial values is often required to avoid convergence issues. As an alternative approach, we propose a block-coordinate descent method that cycles through the considered variables, updating only the parameters related to a given variable in each step. We show that the resulting block update problems can be solved in closed form even when the structural equation model comprises feedback cycles. Furthermore, we give a characterization of the models for which the block-coordinate descent algorithm is well defined, meaning that for generic data and starting values all block optimization problems admit a unique solution. For the characterization, we represent each model by its mixed graph (also known as path diagram), which leads to criteria that can be checked in time that is polynomial in the number of considered variables.
Ann. Statist., Volume 47, Number 2 (2019), 663-690.
Received: October 2016
Revised: May 2017
First available in Project Euclid: 11 January 2019
Permanent link to this document
Digital Object Identifier
Mathematical Reviews number (MathSciNet)
Zentralblatt MATH identifier
Drton, Mathias; Fox, Christopher; Wang, Y. Samuel. Computation of maximum likelihood estimates in cyclic structural equation models. Ann. Statist. 47 (2019), no. 2, 663--690. doi:10.1214/17-AOS1602. https://projecteuclid.org/euclid.aos/1547197234
- Proofs of claims. The supplement provides proofs for claims made in Sections 2, 3 and 4. Specifically, we verify the form of $\det(I-B)$ as claimed in Lemma 1 and derive the likelihood equations with respect to $\Omega$ and $B$. We also verify the claims in Lemmas 4 and 5 which are required for the BCD algorithm described in the constructive proof of Theorem 1. Finally, we verify the claims in Section 4 which characterize graphs for which the BCD algorithm is well defined when initialized generically. In particular, we give a graphical condition and show that it can be checked in time which is a polynomial of the considered variables.