We know that some groups of people are much more likely to be in the child support population than other groups. For example, women are much more likely to be custodial parents than men. These population subgroups can be defined by gender, race or ethnicity, age, and marital status. Large and dramatic differences in the likelihood of being in the child support population exist within each of these groups (see Appendix C). For each group, there are also temporal changes in the likelihood of being a custodial parent. By identifying characteristics associated with being a custodian and by examining temporal trends, we can develop projections of child support populations and understand the forces that are driving changes in the child support population.

In developing projections, we look for stability in historical probabilities of being in the child support population. As shown in Figure 5.1, the percent of adults that were custodial parents rose from around 6 percent in the late 1980s to 7.5 percent by 1995, and, though declining, has remained above 7 percent since 1993. This temporal variation in the probability of being a custodial parent is dwarfed by the variation associated with other characteristics, such as gender, race, age, and marital status. For the purposes of projecting the number of custodial parents in the future, this relatively small temporal variation is good. The much larger variation associated with other factors means that if we can project the population by age, race, gender, and marital status, we can produce reasonable projections of the number of custodial parents.

Because no single data set provides all the information necessary, we used various data sets and methods to develop projections of child support populations. In brief, our method relies on Census Bureau projections of the population by marital status, age, and gender for future populations of the United States. Then, using historic trends in the probability of being in the child support population generated primarily from Current Population Surveys, we forecast and apply those probabilities to the Census Bureau population projections. We disaggregate the trends and hence the projections by race, ethnicity, age, gender, and marital status. We consider the sensitivity of the projections by applying various assumptions and alternative specifications of the projection model. Below, we discuss our approach in more detail.

Our first step is to develop population projections for our various population groups. While the United States Census Bureau provides estimates and projections of the nation’s population by marital status, it does not include race and ethnicity detail. Because the likelihood of being in the child support population varies considerably by race and ethnicity, it is necessary to disaggregate the Census Bureau’s marital status projections into race and ethnic categories. To do so, we use a three-step process. First, we use historical trends in race and ethnic proportions by marital status to project those proportions to 2004 and 2009. Specifically, within distinct marital statuses, we use logistic regression models to identify time trends in the likelihood of belonging to certain race and ethnic groups by age. We extend these trends to 2004 and 2009 by using linear extrapolations. Second, we apply the projected race and ethnic proportions to the Census Bureau’s projections by marital status and age to obtain population projections. Finally, we adjust those projections so that they are consistent with the latest Census Bureau national population projections by age and race/ethnicity. The Census Bureau projections by marital status are relatively old. We evaluated the projections in light of recent estimates, and adjusted our race and ethnic totals to be consistent with the Bureau’s more recent population projections by race and ethnicity (those recent projections do not include marital status). We also discussed with Census Bureau demographers some early results from a new and as of yet unreleased set of population projections by marital status, and were assured that the changes are minimal for our projection horizon (2009).

Our second step involves projecting future likelihoods of being in the child support population. To do so, we analyze past trends in the probability of being a custodial parent and in the probability of being a nonparent custodian. Projections of noncustodial parents and children eligible for child support are derived from the projections of custodial parents and nonparent custodians. We use regression models to identify trends in custodial status among people aged 15 and over. Specifically, using the March Current Population Surveys from 1988 through 1999 and the April Child Support supplements of the Current Population Survey, we develop logistic regression models of the probability of being a custodian. Separate models are developed for custodial parents and nonparent custodians, each of four race and ethnic groups, three marital statuses, and for men and women. This procedure takes advantage of the large variations in the likelihood of being a custodian according to race and ethnicity, gender, marital status, and age.^[11] The regression results provide us with twelve-year trends (1988 through 1999) in the probability of being a custodian. These trends are then used to project future probabilities of being a custodian. Two sets of projections are developed. One set holds the probabilities constant throughout the projection horizon, and the other continues the trends observed during the past five years into the next ten years. These two sets of probability projections for each race and ethnic group, both genders, and each marital status, are then applied to the projected base populations to derive two sets of projections of custodians. The final series of projections then averages these two sets of projections.

To project characteristics not associated with our population groups, we simply use historic trends. In most cases, we use ten-year average annual changes to project changes to 2004 and 2009. One important exception is in the case of public assistance utilization rates among child support populations. As discussed in more detail earlier in the report, we hold those rates constant at 1999 levels.

Finally, we conduct sensitivity tests to evaluate how changes in assumptions and the model would alter the final projections. In general, we find that the projections of custodial parents are less sensitive to changes in assumptions and modeling approach than projections of nonparent custodians. In most cases, the model allowing probabilities to vary over time results in a higher child support population than the model which holds probabilities constant. For custodial parents, the difference in the two sets of projections is less than 5 percent in 2009. For nonparent custodians, the difference is less than 10 percent in 2009. Still, these differences are not insignificant, especially considering the relatively slow growth of the child support population in the United States. They suggest that changes in nonmarital childbearing and divorce could have substantial impacts on future child support populations.