Title: Improving The Success Rate of Optimization Algorithms in
Psychometric Software

Author: Yves Rosseel

Affiliation: Ghent University

Abstract:
Psychometric software often relies on optimization algorithms. In R, the stats
package provides several functions for unconstrained multidimensional
optimization (eg. nlm(), optim() and nlminb()). Often, these optimizers work
well in most cases, but every now and then, they fail to find a solution
(assuming a feasible solution exists). 

In this presentation, I will explain a number of tricks that can be used to
improve the success rate of optimization, when used for estimating the
parameters of a psychometric model. First, I will discuss the use of
theoretical parameter bounds, and how they can be used to stabilize
optimization. Second, I will explain the importance of parameter scaling. And
finally, it will be shown how linear equality constraints can be handled in an
effective way. As a running example, I will use maximum likelihood estimation
in a one-factor CFA model.