Official poverty statistics incorporate the definition of a family that is used in the CPS, and we apply this same definition to compare estimates of income relative to poverty across surveys. A family in the CPS consists of two or more persons living in the same household and related by blood, marriage, or adoption. A CPS family does not include unmarried partners or foster children, but such persons are included in the family definitions of some of the other surveys.15
For two of the eight surveys—NHIS and PSID—we created CPS families within a subset of families that reflected a broader family concept. In each case the family members were reassembled into two or more CPS families, and the income of the original family was apportioned among the new families. These procedures are detailed below. A third survey, MEPS, uses both the CPS family concept and a broader family concept (the same as NHIS), and both are coded on the public use file. For MEPS, then, it was not necessary to create CPS families from more inclusive families; we could use the family data coded on the file.
A fourth survey, HRS, also includes unmarried partners as members of the same family. This affects the family income variable on the RAND-HRS file, which we elected to use for our comparative analysis. Unmarried partners are much less common in the older population that the HRS represents than in the general population.16 Furthermore, our comparative analysis with the HRS data was designed to be much more limited than the analysis involving the general population surveys. For these reasons and because the RAND file lacked suitable personal income variables on which to base a decomposition, we elected to proceed with the HRS analysis without attempting to separate unmarried partners from the family.
Creating CPS Families in NHIS
The family is the basic data collection unit in NHIS. A family respondent provides much of the information obtained from the family, including the family’s total income for the prior calendar year. The concept of family used in NHIS is more inclusive than the CPS family concept. The NHIS family encompasses unmarried partners of the reference person or (in a few cases) of a child or parent of the reference person, whereas the CPS would treat the partner as an unrelated individual. Most commonly, an NHIS family that departs from the CPS family concept will include just a reference person and the reference person’s partner. Most of the rest include children of the reference person or partner but no other adults.17 In addition to including unmarried partners, the NHIS family also includes foster children, who would be treated as unrelated individuals in the CPS, regardless of their age. In all, an estimated 5.8 million or 4.9 percent of NHIS families included unmarried partners or foster children. We designated these “non-CPS” families.18 In order to generate income and poverty statistics from the NHIS that were comparable to the CPS and the other surveys, it was necessary to break up these non-CPS families to form new families that were consistent with the CPS family concept.
Operationally, we achieved this as follows. First, we created two new families (or, in four cases, we created three new families) from each non-CPS family. An unmarried partner of the reference person was assigned to one new family along with anyone identified as that partner’s child. The remaining family members were assigned to the second new family. Some 100,000 foster children age 15 or over were assigned to new, one-person families. Some 200,000 foster children under age 15 were dropped from the sample, as was done for unrelated children under 15 in all surveys. For the families from which these children were dropped, family size was reduced to calculate poverty status, but family income was not changed.
Next, the total family income of each non-CPS family had to be distributed among its subsidiary CPS families. Because there was any number of ways to do this, we elected to apply two alternative algorithms, described in some detail below, in order to determine the possible range of impacts on the poverty count. One algorithm, yielding a lower bound, would distribute the family income in a manner that would produce the fewest number of poor persons. The second algorithm, yielding an upper bound, would distribute the family income in a manner that would produce the most poor persons. We designed and applied the two algorithms and determined that the range between their additions to the poverty count was 430,000 or just 0.15 percent or total persons. Given the small magnitude of the range, we decided to use the average of the lower and upper bounds for each family as a point estimate. That is, for each new family, we calculated two alternative family incomes and assigned the average.
In applying this approach, we made use of personal earnings, which was reported, potentially, for each person 18 and older. We calculated the sum of personal earnings over all members of the NHIS family, calling it family earnings, and compared the result to the total family income. Three scenarios were possible: (1) family earnings and total family income were identical, (2) family earnings exceeded total family income, or (3) total family income was greater than family earnings. What we did next depended on which scenario applied.
If family earnings and family income were equal, then no additional distribution of income was necessary. We assigned each person the amount of his or her own personal earnings and then summed these amounts over the members of each subsidiary CPS family to obtain CPS family incomes that summed to the NHIS family income. The lower and upper bounds were identical.
If NHIS family income was less than NHIS family earnings, we multiplied each person's earnings by the ratio of family income to family earnings. This was done to maintain the original family’s total income (and aggregate family income in the population). This reduced the sum of earnings over all NHIS family members to the amount of total family income. We then summed the reduced earnings over the members of each subsidiary CPS family in order to obtain a family income for each CPS family. Here, too, the lower and upper bounds were identical.
If total family income exceeded family earnings, and the earnings were not zero, we calculated the excess of family income over family earnings and then distributed the excess among the subsidiary CPS families in two alternative ways, representing the lower and upper bounds. With either alternative, each person started with his or her full earnings. For the lower-bound estimate, we assigned the excess family income to the adult with the lowest earnings.19 For the upper-bound estimate, we assigned the excess family income to the adult with the highest earnings. Incomes were then aggregated over the members of each CPS family within the larger NHIS family to create both lower- and upper-bound estimates of family income for each CPS family. The average of the two estimates for each CPS family was then assigned as the family’s income.
For the small number of families (under 400,000 or less than a third of a percent of all families) with no NHIS family earnings, the NHIS family income was apportioned among adults as follows. For the lower bound, we split the family income equally among the adults. For the upper bound, we assigned twice as much income to each adult male as to each adult female, approximating the typical ratio of Social Security benefits between husband and wife, where the spousal benefit is 50 percent of the retiree’s benefit. If the adults were the same sex and there were only two, we assigned two-thirds of the income to the older adult, with the other adult receiving one-third. If there were more than two adults, we assigned twice as much income to the oldest adult as to the rest. As above, we then aggregated each alternative set of incomes over the members of each CPS family to create both lower- and upper-bound estimates of family income for each CPS family. We assigned the average of the two estimates for each CPS family as the family’s income.
Poverty thresholds for all of the new CPS families were determined based on the new family size, number of related children under 18, and whether the family included anyone 65 or older. Estimates of the impact of using the NHIS family concept to assign poverty status are reported in Chapter IV.
Creating CPS Families in the PSID
Like the NHIS, the PSID includes unmarried partners in the same family, except that it does so only for partners of the opposite sex, and in husband-wife or unmarried-partner families the male is always identified as the family head. Relatives of both partners living in the same household are included as well, as are foster children and, in some circumstances, other persons identified as non-relatives of the family head.20 Another departure from the CPS family definition involves families that separated but later reunited (that is, moved back together). Where the CPS would count these as subfamilies within a single family, the PSID continues to treat them as separate families. The family incomes and poverty thresholds for these previously separated families do not reflect their common family membership.
To create CPS families from PSID families that did not conform to a CPS family definition, we had to separate the unmarried partners and combine the related subfamilies. We also had to divide or combine their family incomes and calculate new poverty thresholds that reflected the membership of each family. In addition, we had to remove foster children and other non-relatives.
Operationally, we achieved this as follows. First, we created two or more new families from each non-CPS family. An unmarried partner of the reference person was assigned to one new family along with anyone identified as the partner’s relative. The remaining family members were assigned to the second new family. Foster children and other non-relatives of the family reference person were dropped from the sample, rather than assigned to separate families, because their records contained no personal income data.
Next, the total family income of each non-CPS family had to be distributed among its subsidiary CPS families. Because of restrictions on the income data available, there was little choice about how to do this. The PSID provides some person-level income data for the family head and wife/partner, but certain other components are shared between them. In addition, incomes for all other family members are combined while Social Security is reported as a single amount for the entire family. For the family head, the income components reported are farm income, labor income from unincorporated businesses, asset income from unincorporated businesses, and labor income from employers. For the wife/partner the components are labor income from unincorporated businesses, asset income from unincorporated businesses, and labor income from employers. We assigned the head’s income to the head and the wife/partner’s income to the partner.
The combined asset income of the head and wife/partner from sources other than their respective unincorporated businesses can be calculated by subtracting their individual incomes, as we have just described them, from an amount identified as the taxable income of head and wife/partner. Their combined transfer income, except for Social Security, is reported in a single variable as well. We divided these two sources evenly between the head and partner.
The total taxable income and transfer income (except for Social Security) of all other family members is reported in two additional fields. If one of the two partners had no family members while the other had at least one, then the partner with the family member received all of the income reported for other family members.21 Otherwise, we divided this additional income in proportion to the number of other family members in each family. Thus if the unmarried partner had one other family member while the family head had two, then the family head received two-thirds of the income recorded for other family members.
Lastly, as we have noted, the combined Social Security income of all family members is reported in a single field. If one and only one partner was 62 or older, we assigned all of the Social Security income to that partner. If both partners or neither partner was 62 or older, we divided the Social Security income evenly between them. This completed the apportionment of total family income between the family of the head and the family of the partner.
Related subfamilies living in the same household but treated as separate families can be identified by fields on their respective records. When combining two or three separate families into a single family, we designated the head of the family with the largest total family income as the head of the combined family.22 If the incomes of the separate families were identical, we designated the head of the family with the smaller (or smallest) family ID as the head of the combined family. We summed the family incomes of the two or three separate families to create a family income for the combined family.
If a family member was present for only part of the income reference year, or if another person who was no longer with the family at the time of the interview was present for part of the income reference year, the poverty threshold for that family will reflect the number of months that those persons were present. Likewise, their incomes will be included in the family’s annual income only for those months that they lived with the family.
To account for part-year family members when separating or combining families, we first determined for every family the difference between the poverty threshold reported on the file and the poverty threshold that we would obtain using the reported family size, the number of related children under 18, and whether the head was 65 or older. We defined this difference as the contribution of part-year members to the family poverty threshold. If we separated the families of a head and partner, we assigned this difference to the family of the head. If we combined two or more related families, we summed the values of this difference over the families. When we determined the poverty threshold for a new CPS family, then, we added the value of this difference to the result. Any income received by part-year family members during their period of co-residence with a PSID family would have been included in one of the components discussed above, so there was no need to estimate it separately.
Comparison of Living Arrangements
Even with the application of a common family concept across the five general population surveys, we find differences in the distribution of living arrangements, which are difficult to explain.
After breaking up the non-CPS families in NHIS, we obtain a total of 123.21 million families, which is about 0.7 million more than the CPS (Table III.3).23 Because NHIS counts college students where they are living at the time of the interview, the difference ought to be even greater. Earlier we attributed a 2.2 million shortfall of persons in the ACS to the exclusion of college dormitories and other non-institutional group quarters from the sample frame. NHIS, on the other hand, includes college dormitories in its frame and should be counting the residents of such facilities as unrelated individuals for nine months out of the year, whereas the CPS counts them as members of their parents’ families. However, even with the splitting of unmarried partners we find 1.7 million fewer adult singles (18 and older) in NHIS than in the CPS. This is offset by 5.3 million more married persons in NHIS than the CPS, yet the numbers of married persons ought to be very similar between the two surveys. We are not able to explain this divergence. Rather, we can only suggest that it may stem from differences in the nonresponse adjustments or, more generally, the weighting procedures applied in the two surveys. For example, the NHIS weights do not incorporate a direct adjustment for nonresponding families in responding households, and we suspect that the missing families are primarily single young adults. Post-stratification to population totals may shift the family composition by compensating for too few young adults.
The similar family counts among the CPS, NHIS and ACS suggest that SIPP, with 120.3 million, is at least 2 million too low while MEPS, at 130.90 million, is more than 8 million too high. The excess families in MEPS are especially baffling, as its sample is drawn from responding families in NHIS. We see that large difference between MEPS and the CPS occur in the number of singles, where MEPS is 3.5 million higher than CPS (and 5.2 million higher than NHIS); the number of married childless persons, where MEPS is 3.2 million higher than CPS (but 0.9 million lower than NHIS); and the number of married persons with children, where MEPS is 3.4 million higher than CPS (but only 0.4 million higher than NHIS). In Chapters IV and V we raise the possibility that post-stratification of the MEPS person weights to the CPS poverty distribution may play a role. Differences in the numbers of families would not exist with the MEPS family weights, which are post-stratified to CPS family counts, but as noted earlier, our analysis requires the estimation of person-level characteristics and, therefore, person weights.
|Millions of Families|
|Millions of Persons|
|Single, 18 or oldera||46.91||47.72||45.24||50.40||45.19|
|Children of single parents||20.00||20.73||21.16||19.98||19.31|
|Married, with children||51.89||50.75||51.73||54.12||53.70|
|Children of married couples||48.89||46.66||48.23||50.18||49.59|
|Percent of the Population|
|Single, 18 or older1||16.6||17.2||16.1||17.8||15.9|
|Children of single parent||7.1||7.5||7.5||7.1||6.8|
|Married, with children||18.4||18.3||18.4||19.1||18.9|
|Children of married couples||17.3||16.8||17.2||17.7||17.5|
Source: Mathematica Policy Research, from tabulations of the 2003 CPS ASEC supplement, the 2002 ACS, the 2001 SIPP panel, the 2002 Full-year Consolidated MEPS-HC, and the 2003 NHIS. Note: In the Census Bureau surveys, families include primary families, nonfamily householders, unrelated subfamilies and unrelated (secondary) individuals. Children are under 18.
These differences among the surveys in their estimates of living arrangements raise an important point. If major household surveys cannot agree on something as fundamental as the number of people living alone or with only non-relatives or the number living with spouses, what does this say about their comparative estimates of more complex phenomena? Furthermore, how do differences in the distribution of living arrangements affect estimates of other characteristics? Standardizing on the distribution of living arrangements was not a part of the design of our study, but differences across surveys may have implications for estimates of the poor, or the uninsured, or other subpopulations of policy interest.