In this section of the paper we develop a new indirect method for measuring the historical prevalence of cohabitation. The new method uses data from the March Current Population Surveys, 1977 to 1997, and applies an adjustment to the traditional POSSLQ measure.4
Adjusted POSSLQ Measure
Cohabiting adults may be "living together" alone, with any combination of other relatives, or with other unrelated adults. To improve upon previous indirect estimates by the Census Bureau, in this paper we introduce an "Adjusted POSSLQ" designation that still restricts unmarried couple households to two unrelated adults, but does not exclude households with multiple related adults. The great majority of these related adults are in fact 15- to 17-year-old children of one of the unrelated adults. Research on cohabiting patterns has demonstrated that post-marital cohabitation is increasingly common (Bumpass and Sweet 1989), and these unions often take place in the presence of children from a previous marriage (McLanahan and Casper 1995). The POSSLQ definition excludes such households, leading to an underestimate of cohabiting couples with older children.
We define Adjusted POSSLQ households as those that meet the following criteria: (1) a reference person (householder); (2) one other adult (age 15+) of the opposite sex who is not in a related subfamily, not a secondary individual in group quarters, and not related to or a foster child of the reference person; (3) no other adults (age 15+) except foster children, children or other relatives of the reference person, or children of unrelated subfamilies.
This definition still excludes households in which a reference person lives with a cohabiting partner and that partner's non-child relatives. However, since CPS only records relationships to householders and family or subfamily reference people, these cases are impossible to distinguish from groups of unrelated individuals. More importantly, this definition includes as couple households those situations in which a householder's relative is living with a nonmarital partner. For example, a householder's daughter and her boyfriend might both be present. In such cases, the Adjusted POSSLQ would correctly identify the household as a couple household, but it would incorrectly designate the mother and the daughter's boyfriend as partners.5 Thus, the adjusted measure probably more correctly estimates the number of unmarried-couple households, but might also introduce some patterned biases with regard to partner characteristics. The traditional POSSLQ definition may avoid this latter problem by excluding all such households.
Table 1 shows the number of cohabiting couples as indicated by the Adjusted POSSLQ and traditional POSSLQ methods for 1977 to 1997. The increase in the POSSLQ couples is the now-familiar story of the rapid and nearly linear increase in unrelated couple households, from less than 1 million in 1977 to more than 4 million in 1997. The Adjusted POSSLQ trend reveals the extent of the traditional POSSLQ undercount of potential partner households based on the exclusion of households with related adults. This undercount has increased from about 129,000 in 1977 to 731,000 by 1997, or from about 13 percent to about 18 percent.
Comparisons With Other Data Sets
In 1997, CPS produced a weighted estimate of 3.1 million unmarried partners, 4.1 million traditional POSSLQs, and 4.9 million Adjusted POSSLQs (Table 2). (To compare how the POSSLQ, Adjusted POSSLQ, and unmarried partner CPS measures treat various household scenarios, refer to Appendix A.) Compared to POSSLQ, the Adjusted POSSLQ increases the false-positive rate for unmarried partners slightly from 39.2 percent to 40.7 percent. However, for the three years 1995-1997, the traditional POSSLQ definition did not identify 16.7 percent of the self-identified unmarried partners, excluded because of the number of adults present in these households. In comparison, the Adjusted POSSLQ measure, which only excludes households with more than two unrelated adults, only missed 4.9 percent of the self-identified partners.
To get a better idea of the accuracy of these CPS estimates, we compare estimates achieved with POSSLQ, Adjusted POSSLQ, and our direct measure, with similar estimates from the NSFH, the NSFG, the SIPP and the Consumer Expenditure Survey (CE). Table 3 presents national estimates of cohabitation rates among unmarried women ages 25-44, by age group, for 1987 and 1995. Bumpass and Lu (1998) compared the 1987 NSFH to the 1995 NSFG to measure trends in cohabitation; we use their estimates to evaluate our alternative historical estimates. To their table of direct estimates we add CPS estimates based on POSSLQ, Adjusted POSSLQ, and self-identified partner rates from the 1995 CPS. We also include an indirect estimate from the Consumer Expenditure Survey and a direct estimate from the SIPP.
Direct estimates from the NSFG and NSFH obtain substantially higher rates of cohabitation than all the CPS measures. In 1995, for example, the NSFG estimate of the percentage of unmarried women who are cohabiting in the 35-39 year-old age group is about 13 percentage points higher than the CPS direct measure. The Adjusted POSSLQ measure -- the highest indirect estimate -- is much closer to the estimates in the other two surveys, but for some age groups even this estimate is up to 7 percentage points lower.6 Note, however, that the Adjusted POSSLQ estimates are closer to the 1987 NSFH estimates than they are to those produced by the NSFG in 1995. The direct estimates from SIPP most closely conform to the Adjusted POSSLQ measure. When comparing SIPP estimates with Adjusted POSSLQ estimates, the largest difference is among unmarried women 25 to 29 -- 4 percentage points.
Our only comparison using indirect estimates is with the CE. In their definition, these indirect estimates most closely resemble the traditional POSSLQ: two unmarried adults (16+) of the opposite sex living in a consumer unit. The results are thus close to the CPS POSSLQ estimates.
There are several ways the direct measures of cohabitation prevalence from other surveys could produce higher estimates than the CPS indirect measures. First, they allow identification of partners other than those of the household reference person. For example, about 3 percent of the cohabitors identified in the first wave of the NSFH were not reference people or partners. Second, they may allow multiple couples per household. Third, both the NSFH and NSFG have multiple questions to identify people who are living together who may not have been identified as partners in the relationship questions; this safeguard might also act to boost rates.
Fourth, in some surveys people may say they are "living together" with someone who in the CPS would not be counted as an official member of the household, or who might also be counted as a member of another household because they also have their own home or apartment. That is, "living together" as a relationship state does not necessarily correlate perfectly with official definitions of household membership. Thus, differences in the construction of the household roster and who is or is not considered to be a member may be contributing to some of these differences. For instance, the NSFH roster includes everyone who stays at the house "half the time or more," and NSFG respondents are asked to define relationships with those "people who live and sleep here most of the time." In CPS, the rules for household membership are much more strict. Instructions to CPS interviewers state that a household member "ordinarily stays here all the time," and specifically excludes those who maintain a residence elsewhere (including students).
Fifth, the topic of the survey may influence the identification of cohabitors. For example, more cohabitors might be identified in a survey such as the NSFH whose primary focus is families, or in the NSFG whose primary focus is women's fertility, than in a labor force survey such as the CPS. Sixth, the NSFG only surveys females. To the extent that women are more likely to say they are in a committed relationship (e.g., cohabiting), the NSFG's survey design would identify more cohabitors than a survey such as the CPS which collects information from any knowledgeable respondent.
On the other hand, some have argued that direct measures may undercount cohabitors if people are reluctant to describe cohabiting relationships or think they are not of concern to interviewers. However, to the extent that interviews provide leading phrases to put respondents at ease, undercounting for desirability purposes would be reduced.
Assessing the New Indirect Estimates
In addition to comparing the Adjusted POSSLQ measure to estimates derived from other sources, we also seek to identify potential problems in using the Adjusted POSSLQ to describe trends over time. For the years 1995-1997, the March CPS offers the opportunity to examine the difference between those households identified by the Adjusted POSSLQ measure and those who self-identified as unmarried partners. By modeling this relationship and applying the model to data in previous years, we are able to ascertain if the changing composition of Adjusted POSSLQ households over time differentially affects the validity of the estimates.
The CPS direct measure produces estimates that are lower than the Adjusted POSSLQ measure for the years 1995-1997. To simulate this direct measure for the historical period, we use logistic regression to model self-identified partners as a subset of Adjusted POSSLQ households. We then apply the logistic equation to data from the years 1977-1997 to predict a new estimate of unmarried partner households for those years. In the years 1995-1997, 62 percent of Adjusted POSSLQ households are also self-identified partner households. If the predicted series deviates substantially from the baseline 62 percent, we might have reason to be concerned about using Adjusted POSSLQ as a historical indicator. For example, unmarried couples with children are more likely to be partners than those with no children, and the percentage of couple households with children increased substantially from the 1977 to 1997. Absent other factors, this would suggest that a greater proportion of couple households are unmarried partners today.
Because the Adjusted POSSLQ measure captures more of the self-identified unmarried partners than the POSSLQ measure (Table 2), the Adjusted POSSLQ population appears to be a better universe from which to predict unmarried partners. We use a logistic regression to model unmarried partner households as a subset of Adjusted POSSLQ households for the years 1995-1997 (Table 4). We restrict our analysis to variables that are available in the March CPS for the years 1977-1997, so that the equation may be used to estimate the unmarried partner population for previous years.7 Variable specifications were chosen based on observed bivariate relationships; in the final model some variables no longer have significant effects.8
Figure 1 shows the number of traditional POSSLQ and Adjusted POSSLQ households for 1977-1997, and the number of unmarried partner households for 1995-1997. The figure also includes two estimates for the number of unmarried partners, one based on the assumption that the unmarried partner population has remained a constant 62 percent of the size of the Adjusted POSSLQ population, and the other the predicted population from the logistic equation. The prediction model produces estimates somewhat higher than the .62 constant model from 1978 to 1988, and slightly lower than the constant model for the years 1989 to 1997. This suggests that the Adjusted POSSLQ measure is slightly more likely to overestimate the actual unmarried partner population in more recent years, but the predicted trend does not deviate substantially from the constant series. Therefore, we conclude the Adjusted POSSLQ is an acceptable measure for historical trends in the prevalence of cohabitation.
In this section, we have compared the direct and indirect CPS measures of cohabitation with those obtained from other data sets. We have also used direct estimates to evaluate whether our new Adjusted POSSLQ measure yields valid estimates over time. These analyses have been instructive in that they have exposed the limitations of various approaches to measuring cohabitation. Given the advantages and limitations, which series of estimates is better for use in describing historical trends and which is better for examining the changing composition of cohabitors? The Adjusted POSSLQ measure produces estimates that appear closer to the intended definition of cohabitation than the traditional POSSLQ measure, and resolves some of traditional POSSLQ's known shortcomings. Because of this, and because the Adjusted POSSLQ appears to hold up well over time, we conclude that the new Adjusted POSSLQ measure is a better measure for monitoring aggregate historical trends.