State Estimates of Uninsured Children, January 1998. Final Report.. C. Methodology


The estimates presented in this report were derived from the March 1998 CPS public use file after making a number of enhancements discussed below and in the Technical Appendix. The CPS is a monthly sample survey of 50,000 households in the United States. It serves as the source of the official monthly unemployment estimates for the nation and the 50 states and the District of Columbia. Interviews are conducted with all civilian non institutionalized persons age 15 and older in the sample households. Each March, the Census Bureau includes a supplemental questionnaire in the CPS interview. The data collected in this questionnaire are the source of the official estimates of the incidence of poverty in the United States. In addition, the March supplement collects information on the health insurance coverage of all household members (including children). These data have become the source of the most widely-cited estimates of the number of uninsured children in the United States each year.

While the CPS sample was designed to support state-level estimates of unemployment rates and other labor force statistics, the sample sizes for most states are inadequate for satisfactorily precise monthly estimates that use only the survey data for that month.2 Thus, even if we are producing just a single estimate for each state--such as the number of uninsured children--that estimate will generally be imprecise due to the high sampling error associated with small samples. This means that we will be very uncertain about the true number of uninsured children and able to say for a typical state only that the number falls within a wide range, a range that is too wide to provide useful guidance for developing policy or administering a program.

This problem is even more severe when we must produce many estimates for each state, such as a table showing the distribution of uninsured children across poverty and age categories. Then, the already small sample for a state must be spread across the many cells of the table. It is not unusual for the available sample to have no observations in some cells even though the state population obviously has uninsured children with the characteristics defined by those cells.

When large samples are not available, a standard approach to improving the precision of sample estimates is to borrow strength.3 This entails the development of statistical models that allow us to derive, say, a 1998 estimate of uninsured children in Virginia using not only the 1998 data for Virginia from the main sample survey database but also data from other states, earlier years, and auxiliary sources, such as administrative records. Knowing something about the numbers of uninsured children in Virginia in 1996 and 1997 and the numbers of uninsured children in other states with economic and demographic conditions similar to Virginia's, we would generally be able to reduce substantially our uncertainty about how many uninsured children lived in Virginia in 1998.

To derive most of the estimates presented in this report, we have used a method for borrowing strength that is described in detail by Schirm and Zaslavsky (1997). With this method, we have reweighted the March 1998 CPS database to produce 51 sets of weights--one set for each state and the District of Columbia. We use the Virginia weights to derive estimates for Virginia. Each of the roughly 50,000 households in the CPS database--regardless of state of residence--gets a Virginia weight, greatly increasing the size of the sample from which to obtain estimates for Virginia. How much Virginia weight a household gets and, therefore, its relative contribution to estimates for Virginia depends on the household's characteristics. If a household from, say, Montana has characteristics that would make it unusual were it in Virginia, as opposed to other states, it receives a relatively small--probably negligible--Virginia weight. If, instead, a household of that type would be more common in Virginia than in other states, it receives a relatively large Virginia weight.

The Virginia weight assigned to a household with a particular set of characteristics depends on the aggregate characteristics of Virginia. Specifically, Virginia weights are controlled so that totals derived for Virginia using Virginia weights equal specified values. For example, the weights might be controlled so as to reproduce specified totals of children, children below 100 percent of poverty, and uninsured children. By controlling the weights, we ensure that the entire database--when weighted according to the Virginia weights--looks like Virginia in terms of totals that are relevant to ascertaining the patterns of insurance coverage among children.4

In reweighting the March 1998 CPS database, we used controls reflecting the age and racial/ethnic structure of each state's child population, the distribution of children across poverty categories, and the numbers of uninsured children (by poverty category). The full list of totals to which we controlled weights appears in the Technical Appendix.

Reweighting a database as we have described can substantially improve the precision of estimates because the samples used to derive the estimates are much larger than when we use only the observations from a single state. Using observations from all the states allows us to borrow strength. Although this alone improves precision, we have further improved precision by using administrative estimates or empirical Bayes shrinkage estimates--rather than direct sample estimates --for many of the control totals used in the reweighting.5 The administrative totals, which are population estimates derived from mainly vital records (and decennial census) data, have essentially no sampling error, and the shrinkage totals, which are derived by borrowing strength, are more precise than direct sample estimates. The specific sources of control totals are described in greater detail in the Technical Appendix.6