For each local area and for each item stratum, an estimate of consumer expenditure is needed to weight the market basket of goods and services for which the U.S. Consumer Price Index (CPI) is computed. These expenditure estimates, called cost weights, must be computed every time the CPI is revised. Due to small sample sizes in the Consumer Expenditure Survey, high variability is an inherent problem in producing such localized estimates. To alleviate this problem, the U.S. Bureau of Labor Statistics (BLS) employs composite estimation to reduce the mean squared error of the cost weight estimates at the index area/item stratum level. In this paper we summarize the research conducted at the BLS over the past ten years on different methods of composite estimation, and describe the method used in the CPI's 1998 revision.