Unexpectedly, the estimates of margins of error (MOE) for a small number of 2005-2009 ACS characteristic estimates are larger than the MOEs for the same statistics based on 2007-2009 ACS estimates, and to a lesser extent the MOEs for the 2009 estimates at the national and state levels. For example, we are seeing the 5-year MOE for the estimate of total number of households is over twice the size of its 3-year MOE and is about one and a half times the size of its 1-year MOE. This is unexpected but does not necessarily reflect an error in processing. In theory, the 5-year MOEs should be smaller than both because the estimates are based on larger sample sizes. However, the estimates of MOE depend on other factors in addition to sample size.

The margins of error are estimated using replicate weights. Looking at various stages of the weighting process, we see a substantial increase in some multi-year MOEs occurring at a step known as g-weighting, a method that is not used for the 1-year estimates. This weighting step adjusts the housing unit population counts by demographics using auxiliary data from administrative data sources in an attempt to reduce both the bias and MOE of the estimates. For the 3-year estimates and for the 5-year estimates for small areas, the g-weighting adjustment generally reduces both the bias and estimated MOE. However for the 5-year estimates for large areas, g-weighting occasionally increases the estimated MOEs. That is, the impact of the g-weighting is different at the national and state levels than it is for small areas, such as census tracts. At this point, there is no evidence that the MOE estimates were computed incorrectly (i.e. not according to specifications) or that a processing error was made. Furthermore, the MOEs for national and state estimates are already so low that a modest increase will typically not affect inferences made with those estimates.

We have concluded that this unexpected phenomenon does not affect the 5-year estimates in any important way. The g-weighting adjustment is helping to reduce MOEs of the 5-year estimates at lower levels of geography (counties, places, and tracts) as expected. Further investigation of the properties of g-weighting for large area estimates is planned.