State Model Changes
Some changes were made to the sampling error models used to develop smoothed sampling error variances. These changes returned us to using the same approach we used prior to 2000. This year the sampling error models were fitted to sampling covariance matrices estimated for income years (IYs) 2000-2003.
Last year's models used different design effect parameters for each year out of concern that some changes in 2002 to the CPS ASEC sample weighting scheme could affect the design effects for the variances. Fitting models with different design effect parameters for each year yielded estimates of these parameters that were, in fact, rather similar for each year, and statistical comparisons of these models with models that used common design effect parameters each year slightly favored the latter. So this year we returned to using a common design effect parameter for all years.
For the last two years we removed state random effects from our sampling error models out of concerns that we might not be able to estimate these effectively from sampling error covariance matrices estimated for just two or three years of data. (The state random effects seek to account for any given state having systematic over- or under-estimation of its sampling error variance by the fitted generalized variance function.) Having an additional year of data in our sampling error covariance matrices gave us more confidence that we could estimate the state random effects. Also, statistical comparisons of models with and without the state random effects strongly favored the models with these effects. So we returned to including state random effects in our sampling error models and using the estimates of these effects to adjust the fitted generalized variance function values for each state.
As was done last year, for 2003 we made multiplicative adjustments to the poverty ratio variances obtained from the fitted sampling error models to approximately reflect the effects of the changes made starting in 2002 to the CPS ASEC sample weighting scheme. The effects of these variance adjustments were small last year, and even smaller for 2003, with the largest effect being a decrease in the 18-64 fitted sampling error variances of about five percent. Such adjustments were not made to the sampling variances for the median income estimates because the approach used to obtain approximate variances (which used a linearization approach) was thought to be less appropriate for the median income estimates.
For further discussion of the sampling error models see "Sampling Error Modeling of Poverty and Income Statistics for States," (Otto and Bell, 1995).
County Model Changes
There were no methodology changes to the county models.