The Bureau of Labor Statistics is conducting research on the feasibility of producing industry employment estimates by size of firm using the Current Employment Statistics (CES) survey, a monthly survey of business payrolls. The Office of Management and Budget has designated 12 size classes to serve as the basis for official estimates produced by federal statistical agencies, but a smaller number of categories must be used for CES estimates in order to reduce sampling error. This paper analyzes the problem of collapsing the 12 classes into a smaller number of categories for use in constructing CES estimates. We develop a methodology based on K-means cluster analysis modified to require that clusters retain the ordering of classes. We also require that each cluster has sufficient employment share and adequate sample in each industry. We use 4 characteristics of size classes in the cluster analysis: seasonality of employment, longterm trend in employment, cyclicality of employment, and share of employment related to business births and deaths. We also consider size categories used in related data and reclassifications of firms over time across size categories.