Abstract
The rolling sample methodology of the American Community Survey leads to Multi-Year Estimates that measure aggregate activity over one, three, or five years. This paper introduces a novel, non-model-based method for quantifying the impact of viewing multi-year estimates as functions of single-year estimates belonging to the same time span. The method is based on examining the changes to confidence interval coverage. The interpretation of a multi-year estimate as the simple average of single-year estimates is a viewpoint that underpins the published estimates of sampling variability. Therefore, it is vital to ascertain the extent to which this viewpoint is valid. We apply our new methodology to data from the U.S. Census Bureau's Multi-Year Estimates Study and demonstrate that viewing a multi-year estimate as the simple average of single-year estimates typically results in substantial distortions to coverage; therefore, multi-year estimates should not be interpreted as averages, but merely as period estimates.