Data on employee benefits is one of the key products produced by the National Compensation Survey (NCS). The NCS uses a stratified three-stage sample: areas (PSUs), establishments, and occupations. The weighting procedure includes adjusting sample weights for establishment non-response, occupation non-response, any unusual data collection situations, and post-stratifying to known employment control totals. This paper illustrates how Fay's method of balanced repeated replication (BRR) was applied to estimates generated from this complex sample design to obtain variance estimates for the NCS benefits products. The following topics will be presented: an introduction to the NCS stratified three-stage sample; a description of our utilization of Fay's method, including formation of variance strata, variance PSUs, and replicates; a detailed analysis of the magnitude of estimated variances.