A latent class model was considered, with estimates attained by appling the expectation-maximization (EM) algorithm to the log likelihood function. It was found that the estimates are sensitive to the chosen starting points; remedies were investigated. Results are reported here.