Whist considerable attention has been Paid to determining the number of classes in a latent class analysis less attention has directed at the optimal selection of indicator variables . Indicator selection reduces redundancy and complexity,and can provide a way forward in cases where the number of indicators in large. However, determination of the optimal indicator set and the optimal number of classes is not straightforward, as the two are heavily interrelated. This paper reports on a reformulation and extension of the dean and Raftery algorithm. By treating subset selection as an imposition of sets of constraints on the class membership probability , the BIC (or any other information criterion) becomes informative both for determining the optimal subset selection and for determining the number of classes. The procedure is illustrated by a dataset on the presence or absence of psychiatric symptoms in 30  psychiatric  patients.