PSYCHOMETRIKA--VOL. 50, NO. 2, 229-242
JUNE 1985
NONCONVERGENCE, IMPROPER SOLUTIONS, AND STARTING VALUES
IN LISREL MAXIMUM LIKELIHOOD ESTIMATION
ANNE BOOMSMA
UNIVERSITYOF GRONINGEN
In the framework of a robustness study on maximum likelihood estimation with LISREL
three types of problems are dealt with: nonconvergence,improper solutions, and choice of start-
ing values. The purpose of the paper is to illustrate why and to what extent these problems are of
importance for users of LISREL. The ways in which these issues may affect the design and con-
clusions of robustness research is also discussed.
Key words: maximum likelihood estimation, LISREL, Davidon-Fletcher-Powell, starting values,
nonconvergence,improper solutions, robustness, Monte Carlo, small sample results.
Introduction
In studies on the robustness of maximum likelihood estimation with LISREL, the
effects of small sample size and those of nonnormal, discrete distributions were investi-
gated (Boomsma, 1982, 1983). Three main conclusions from that research were made. (i)
In LISREL modeling it is recommended not to use a sample size smaller than 100. (ii)
LISREL is robust against symmetric, discrete distributions with normal kurtosis, but not
against rather skewed, discrete distributions. (iii) It is not recommended to analyze corre-
lation matrices instead of covariance matrices, because it may have serious effects on the
estimated covariances of the parameter estimates.
The present paper deals with three problems encountered during the work just re-
ferred to: nonconvergence of the iterative maximum likelihood estimation, improper pa-
rameter estimates, and the choice of starting values in the estimation process.
The results discussed below indicate that users of LISREL should be aware of cir-
cumstances in which nonconvergence and improper solutions may occur, and their fre-
quency of occurrence. They may also be interested in the effect of the starting values in
iterative estimation. Some of the materials presented may serve as guidelines for re-
searchers planning Monte Carlo studies. The emphasis in this paper will be on the practi-
cal implications of the findings for users of LISREL, illustrated by results from Monte
Carlo work.
Monte Carlo design. The sampling design for the small sample part of our study is
summarized as follows. In the LISREL framework for each specified model with a known
population parameter vector to, a population covariance matrix ~ is defined. These
matrices ~ are the covariance structures of multivariate normal distributions from which
independent samples were taken of size 25, 50, 100, 200, and 400. For each model and
each sample size N, NRS > 300 samples were taken (NRS = number of replications in
stock). The samples were generated by using subroutine GGNMS from IMSL (1982) on a
CDC Cyber 74/18 and a CDC Cyber 170/760 machine. Thus, for each model and each
sample size, NRS sample covariance matrices S were obtained.
After this sampling process a LISREL analysis was done for each S until NR = 300
samples have led to a solution without numerical difficulties; for all N this number of
Requests for reprints should be sent to A. Boomsma, Vakgroep Statistiek en Meettheorie, Rijksuniversiteit
Groningen, Oude Boteringestraat 23, 9712 GC Groningen, THE NETHERLANDS.
0033- 3123/85/0600- 5010500. 75/0 229
© 1985The Psychometric Society