The estimates from an EM, when it is applied to a large causal model of 10 or more categorical variables, are often subject to the initial values for the estimates. This phenomenon becomes more serious as the model structure becomes more complicated involving more variables. As a measure of compensation for this, it has been recommended in literature that EMs are implemented several times with different sets of initial values to obtain more appropriate estimates. We propose an improved approach for initial values. The main idea is that we use initials that are calibrated to data. A simulation result strongly indicates that the calibrated initials give rise to the estimates that are far closer to the true values than the initials that are not calibrated. (C) 2002 Published by Elsevier Science B.V.