The Current Population Survey, conducted by the Census Bureau for the Bureau of Labor Statistics, provides official labor force estimates for the United States. In the two-stage design, primary sampling units are selected and housing units are selected within those PSUs. A variance is logically the sum of a between-PSU component and a within-PSU component. For national variances, there are enough PSUs to allow estimating total variances using a successive difference replication method. However, for a state, it is necessary to separately estimate the two components. This paper concentrates on within-PSU variance estimation, based on within-PSU replicate assignments by state. The calculated estimates of within-PSU replicate variances are inherently "noisy", so generalized variance functions (GVFs) are modeled over time for employment and unemployment in each state. Innovative changes in methodology and functional form allow many more years of data to be used in developing stable GVFs. Selected GVFs are graphically compared to the underlying within-PSU replicate estimates, and the modeling allows for an interesting assessment of state within-PSU design effects.