Measuring Cohabitation in U.S. National Surveys

Abstract

Cohabitation is one of the fastest growing family forms in the United States. It is widespread and continues to increase but has not been consistently measured across surveys. It is important to track the quality of data on cohabitation because it has implications for research on the correlates and consequences of cohabitation for adults and children. Recent rounds of the Current Population Survey (CPS), National Longitudinal Survey of Adolescent to Adult Health (Add Health), National Longitudinal Survey of Youth (NLSY-97), and National Survey of Family Growth (NSFG) provide an opportunity to contrast estimates of cohabitation status and experience using nationally representative data sets and assess the quality of data on cohabitation in these data sets. Results demonstrated that the surveys provide similar estimates of current cohabitation status, except the CPS resulted in lower estimates. In terms of cohabitation experience (i.e., having ever cohabited), Add Health produced higher estimates, whereas both the NSFG and NLSY-97 produced lower estimates. We documented a strong education gradient across all surveys, with lower levels of current cohabitation and cohabitating experience and with increases in educational attainment. Racial/ethnic differences in cohabitation were inconsistent across surveys. We discuss aspects of sampling and measurement that potentially explain differences in estimates. Our findings have implications not only for survey design but also for the interpretation of results based on these four national surveys.

Introduction

Cohabitation has increased rapidly in the United States, and this growth has been termed a Cohabitation Revolution (Smock and Manning 2010). Today, cohabitation is experienced across the life span, with the majority of young adults spending some time in a cohabiting union (Manning and Stykes 2015) and an increasing share of older Americans living with their cohabiting partner (Brown et al. 2012). The number of cohabiting couples is at a historic high point, surpassing 8 million in 2015 (U.S. Census Bureau 2015). The measurement of cohabitation status and experience has implications for demographic research. Demographers have relied on many surveys to assess the correlates associated with transitions into cohabiting unions, the stability of cohabiting unions, childbearing patterns of cohabiting couples, and the well-being of children and adults in cohabitation (e.g., Addo 2014; Brown, et al. 2017; Guzzo 2017; Kennedy and Fitch 2012; Kuo and Raley 2016; Manning 2015; Musick and Michelmore 2015).

Cohabitation is so common that it is now included as a relationship status in nearly all major national surveys. Demographers have long focused on the measurement of cohabitation in the United States (e.g., Bumpass and Lu 2000; Casper and Cohen 2000). Early estimates of cohabitation were based on indirect measurement, such as POSSLQ (persons of the opposite sex sharing living quarters), but more recent surveys have shifted to direct questions establishing cohabitation status. Even though recent rounds of surveys have included direct questions about cohabitation, the wording of questions measuring cohabitation varies. Rarely have researchers contrasted the measurement of different-gender cohabitation across surveys, and a comparison is long overdue (for exceptions, see Bumpass and Lu 2000; Casper and Cohen 2000; Hayford and Morgan 2008). We expect gaps in measurement across surveys to exist, but we also anticipate that the gap has narrowed during a period when cohabitation is quite common and socially accepted.

In this article, we contrast cohabitation status and experience for a comparable cohort of young adults in a specific time period across four national, population-based surveys widely used to examine cohabitation: the Current Population Survey (CPS), National Longitudinal Study of Adolescent to Adult Health (Add Health), National Longitudinal Study of Youth-97 (NLSY-97), and National Survey of Family Growth (NSFG). Given the age and period restrictions of each data collection as well as the constraint of only opposite-sex cohabitation in the NSFG, our comparisons across these major surveys are limited to young adults (26–28 years old) and their different-gender cohabiting relationships. We assess reports of young adult different-gender cohabitation status and experience across these surveys and examine how levels differ according to gender, educational attainment, and race/ethnicity. We argue that researchers should be aware of differences in the measurement of cohabitation across surveys to ensure accurate estimates of cohabitation and best assess its correlates and implications.

Background

Cohabitation was recognized as an increasingly common and important relationship status in the 1970s in the United States (e.g., Glick and Norton 1977; Macklin 1978). One strategy in early efforts to study cohabiting couples was to base work on small-scale studies and/or convenience-based samples with dedicated measurement of cohabitation status (Blumstein and Schwartz 1983; Macklin 1978). Another strategy was to use population-based surveys to infer cohabitation status based on age, marital status, gender, and family composition. Initial efforts include work using the 1975 Current Population Survey (CPS) data (Glick and Spanier 1980). The POSSLQ acronym referenced these indirect measurement strategies. Beginning in the late 1970s, the U.S. Census Bureau classified POSSLQs as households with unmarried respondents 16 or older of the opposite sex and no other adults (see Casper and Cohen 2000). In their review of this measure, Casper and Cohen (2000:237) stated, “The definition thus misses cohabitors who share households with other adults, and at the same time includes adults who live together without being couples, such as college roommates.” A variety of strategies followed to establish adjusted indicators of cohabitation, such as limiting age differences of partners and/or allowing the inclusion of related adults (Baughman et al. 2002; Casper and Cohen 2000; Fitch et al. 2005; Moffitt et al. 1998; Winkler 1993).

The research community and federal survey sponsors kept pace with the changing relationship landscape by including measures of cohabitation in national population-based surveys in subsequent years. In the late 1980s, national surveys started to add direct questions (NSFG) and incorporate further detail including cohabitation histories (National Survey of Families and Households (NSFH)). “Unmarried partner” was added as a relationship status on household rosters in the 1990 Decennial Census and 1995 CPS. In the mid-1990s, several additional population-based surveys followed suit, including cohabitation status (Survey of Income and Program Participation (SIPP), NLSY-97) and asking questions to establish full cohabitation histories (NSFG, 1995).

Comparisons of direct and indirect measurement of cohabitation within surveys revealed several types of error. Analysis of the CPS, decennial census, and SIPP showed that indirect measurement (e.g., POSSLQ) resulted in differential counts of cohabiting couples than direct measurement based on household rosters (Baughman et al. 2002; Casper and Cohen 2000; Fitch et al. 2005; Manning 1995). Both measurement strategies were recognized as introducing bias and error.

The first studies comparing levels of cohabitation across surveys were published more than 15 years ago (referencing cohabitation levels more than 20 years ago). Two key studies published in 2000 benchmarked levels of cohabitation across surveys (Bumpass and Lu 2000; Casper and Cohen 2000). These studies compared cohabitation levels 1987/1988 using the NSFH, NSFG, and CPS (Bumpass and Lu 2000; Casper and Cohen 2000). Levels of cohabitation based on direct measures in the family surveys, NSFG and NSFH, were quite similar but much higher than levels based on the indirect methods in the CPS. For 1995, Casper and Cohen (2000) compared cohabitation levels for women across the following national surveys: Consumer Expenditure Survey (CE), CPS, NSFG, and SIPP. They reported that cohabitation levels based on the direct measures in the NSFG were greater for each age group than levels based on the direct measures in the SIPP and CPS and the indirect methods in the CE. The consensus was to use direct measures and to not rely on household rosters to measure cohabitation. However, there has been no update of measurement of cohabitation across surveys, which is problematic because these studies documented levels of cohabitation in 1987/1988 and 1995, when the number and share of young adults who ever cohabited was 50 % lower than today (Manning and Stykes 2015).

Nonetheless, attention to the measurement of cohabitation has continued. A significant recent development in the measurement of cohabitation is the use of partner pointers in the CPS to more effectively identify cohabiting partners (Kreider 2008). Rather than relying solely on reports of the household head’s relationship to each household member to identify cohabiting couples, the CPS began using questions in 2007 to identify all members of the household who were cohabiting. The head of household was asked whether each unrelated and unmarried member of the household had a “boyfriend, girlfriend, or partner in the household.” Cohabiting partners could be linked using line numbers from the household roster. This direct strategy has identified a sample of cohabitors who differ on a range of sociodemographic indicators from the sample identified through roster methodology (Kennedy and Fitch 2012; Kreider 2008). Important benefits of this approach are identifying cohabiting partners who do not identify as unmarried partner in the household roster and cohabiting couples in which neither is head of household.

Although prior comparative studies have focused on point estimates of cohabitation, it is important to include cohabitation experience indicators. The short duration of cohabiting unions (18 months on average) means that point estimates are not well equipped to capture whether individuals ever have cohabited. The share of Americans currently cohabiting is quite low in contrast to the share who have ever cohabited. On one hand, there may be greater variation across surveys in the point estimates than ever-cohabited indicators because the boundaries between cohabitation and singlehood are sometimes blurry. This is evident from studies that documented the gradual processes of moving in/out, along with high rates of relationship churning (breaking up and getting back together) (Avellar and Smock 2005; Halpern-Meekin et al. 2012; Knab and McLanahan 2007; Manning and Smock 2005; Pollard and Harris 2007; Sassler 2004). As a result, there may be more consensus across surveys in whether young adults have ever cohabited than whether they are cohabiting at the time of interview. On the other hand, point estimates across surveys may be more similar because these indicators avoid retrospective bias, especially over long intervals (10 years or more) (Hayford and Morgan 2008; Teitler et al. 2006).

Researchers have moved beyond traditional approaches to pioneer alternative ways to assess data quality by relying on in-depth interviews as well as surveys employing several indicators and multiple reporters of cohabitation. In-depth interviews showcase potential problems using the term unmarried partner (Manning and Smock 2005). As a result, several surveys have included the term boyfriend/girlfriend to the list of relationship options. The specific wording of questions to measure cohabitation in surveys matters; Pollard and Harris (2007) found that referencing cohabitation as “marriage-like” leads to lower estimates of cohabitation than items asking about living together (Pollard and Harris 2007).

Another data-quality issue has been establishing the start and end dates of cohabitation. Qualitative data collections demonstrate that it is not a simple transition into and out of cohabiting unions (Manning and Smock 2005; Sassler 2004). Transitions to marriage have an obvious start date that is celebrated annually with anniversaries (Wu et al. 2011). For instance, using data from the Fragile Families and Child Wellbeing Study (hereafter, Fragile Families), Knab (2005) introduced the idea of “cohabitation as a fuzzy concept” by showing that one in six mothers are cohabiting part-time (fewer than 6 or 7 nights). Indeed, many young adult couples in the third wave of the Add Health spent the night together on a regular basis (part-time cohabitation) prior to occupying in single residence (Pollard and Harris 2007). This ambiguity in defining cohabitation has been further illustrated when members of the same family do not always report cohabitation in the same manner. For example, in the Add Health, adolescents and their parents do not always agree about parental cohabitation status (Brown and Manning 2009), and unmarried parents in the Fragile Families disagree about their cohabitation status at the time of birth (Knab and McLanahan 2007). Relatedly, married couples do not share similar reports of premarital cohabitation, with roughly 1 in 10 married couples in the NSFH disagreeing about their cohabitation experiences (Thomson and Colella 1992) and nearly one-half of married couples who cohabited premaritally in a 2010 Internet survey differing in their reporting of the timing and/or duration of cohabitation (Halpern-Meekin and Tach 2013).

A data-quality issue specific to the measurement of cohabitation experience has been questions about the reliance on retrospective reports. Teitler et al. (2006) reported that one in eight mothers were inconsistent in their reporting of cohabitation status at initial interview (time of child’s birth) and retrospective report of cohabitation status at one year after the birth. Similarly, Hayford and Morgan (2008) determined that retrospective reporting of cohabitation experience results in lower levels of reporting cohabitation experience than more contemporaneous reporting.

Current Investigation

Cohabitation continues to garner extensive research attention, but no recent study has compared measurement of cohabitation across any of these recent surveys used to study cohabitation: Add Health, CPS, NLSY-97, and NSFG. Each data set offers a different lens on the measurement of cohabitation, based on sampling, interview mode, frequency, questionnaire design, and wording. The Add Health and NLSY-97 allow two estimates of cohabitation based on rosters and direct survey questions. We generate estimates of early adult cohabitation experience for a comparable cohort of early millennials (born 1979–1982). We focus on young adults who were ages 26–28 in 2007–2008: all four surveys cover this age group during this period.

We address three research questions. First, we examine how levels of current cohabitation status and cohabitation experience compare (the CPS cannot be used to assess cohabitation experience). We expect estimates to differ beyond ordinary random variation across these surveys. Second, we compare levels of cohabitation across data sets according to gender, age, education attainment, and race/ethnicity. Given reported differences in cohabitation experience by education, race/ethnicity, and gender (Hemez et al. 2017; Kennedy and Fitch 2012), we present results separately for each subgroup. Further, because some sociodemographic groups more often report potentially part-time or in-flux relationships (e.g., blacks, lower education and income, younger ages, men), there may be subgroup variation in the consistency across data sets in the reporting of cohabitation (Knab 2005; Knab and McLanahan 2007; Nepomnyaschy and Teitler 2013; Pollard and Harris 2007; Teitler et al. 2006; Vennum et al. 2014). Third, using a standardization technique, we investigate whether a potential explanation for variation in the reporting of cohabitation across surveys is the difference in the sociodemographic characteristics of each of the analytic samples. This is important because estimates of cohabitation may differ across surveys depending on the extent to which they over- or underrepresent various subgroups.

Data and Methods

We draw on four nationally representative population-based data collections: Add Health, CPS, NLSY-97, and NSFG. To ensure comparability across surveys, we restrict each data set to a specific birth cohort (those born in 1979–1982) and period (2007–2008). We also limit our sample to respondents who are not institutionalized (i.e., in military housing, prison, or jail) because the CPS and the NSFG exclude these populations from their sampling frame. Thus, this sample reflects the experiences of a cohort in young adulthood (aged 26–28). The wording of the NSFG question on cohabitation is limited to opposite-sex cohabitation, requiring a focus on different-gender cohabitation status and experience across all surveys. All data are weighted based on the recommendations provided in the user’s guides by the respective data providers. We provide an overview of each survey, both in this section and in Table 1, with a focusing on the following: sampling; interview mode and frequency; sample sizes; and question design and wording. We provide extensive detail on these features of each survey in the online appendix.

Sampling

A relevant difference across surveys is the sampling unit. Respondents in the NSFG, Add Health, and NLSY-97 provided information on their own behaviors, so all cohabitations were identified directly from the cohabitor. The CPS is a household-based survey. In the CPS, respondents were knowledgeable household heads (aged 15 or older) who provided information for all individuals currently living in the household. If the household head had a cohabiting partner, cohabitation was identified through direct reporting. However, if the household head reported that someone else living in the household had a cohabiting partner, cohabitation was identified through proxy reporting (described later).

The sampling frame differs somewhat across surveys. The Add Health data collection is a school-based sample that required respondents were enrolled in 7th–12th grade during the 1994–1995 school year. The NSFG, NLSY-97, and CPS are population-based surveys that did not require enrollment in school to be included in the survey. The Add Health sampling strategy may result in a more highly educated set of respondents (i.e., respondents who had not dropped out of school).

Interview Mode and Frequency

The respondents in the NSFG and Add Health were interviewed in person; however, Add Health used audio computer-assisted self-interviews (ACASI) to obtain information on relationship histories and other sensitive information. The NLSY-97 respondents were largely interviewed in person (in 2007, about 13 % were interviewed via phone). CPS respondents were interviewed using a combination of phone and face-to-face interviews. Interviewer-based strategies provide opportunities for respondents to query about definitions of terms, such as unmarried partner.

The Add Health and NSLY-97 are longitudinal data collections, and the CPS and NSFG are a cross-sectional design. The initial Add Health interview occurred when respondents were in grades 7–12 in 1994–1995. We limit our analyses to those who participated in the initial interview and Wave 4 (2007–2008). The NSLY-97 interviewed respondents in 1997 when they were aged 12–16 and has conducted yearly interviews until 2010; after 2011, interviews occurred every two years. The NSFG segmented the 2006–2010 interview into quarters based on interview dates; we select respondents interviewed in the June 2006 to December 2008 period (quarters 1–10). We select CPS respondents who were interviewed in 2007 or 2008. The analyses are limited to respondents who met the age restrictions during the 2007–2008 period.

Sample Sizes

The sample sizes vary considerably. The cohort studies interviewed 15,701 Add Health and 8,984 NLSY-97 respondents; the NSFG consists of 22,682 respondents aged 15–44; and CPS is based on 75,872 respondents over age 15. As a result, the analytic sample consists of 26- to 28-year-olds in 2007–2008 range: 6,264 in the Add Health; 4,349 in the NLSY-97; 1,518 in the NSFG; and 11,543 in the CPS. The Add Health and NLSY-97 analytic samples are subject to attrition (discussed later).

Questionnaire Design and Wording

Each survey used unique questions to identify cohabitation, which might influence cohabitation estimates (detailed description provided in the online appendix). To summarize, the CPS asks about a “partner” as well as “girlfriend, boyfriend or partner”; the NSLY-97 queries about living in “marriage-like” sexual relationships; the NSFG specifies a sexual relationship sharing a “usual residence”; and the Add Health refers to sharing a residence with a “romantic or sexual partner.”

Current Cohabitation Status

We adopt and compare two strategies to identify cohabiting couples in the Add Health based on the household roster as well as survey questions in the relationship history section. The Add Health asked the respondent to list all individuals living in their household and established a roster. Add Health then asked, “Is {INITIALS} male or female?” and “What is {INITIALS} relationship to you?” If respondents claimed that the individual is a partner/boyfriend/girlfriend of a different gender, they were coded as currently cohabiting. The second measure relies on information from the relationships section, using ACASI-generated romantic and sexual partner rosters. ACASI maximizes privacy and allows for complicated skip patterns (Paik 2015). The relationships section asked never-married respondents, “How many romantic or sexual partners have you ever lived with for one month or more?” It asked ever-married respondents, “Not counting the (partner/partners) you married, how many other romantic or sexual partners have you ever lived with for one month or more?” Both groups of respondents were instructed, “By ‘lived with’ we mean that neither of you kept a separate residence while you were living together.” Respondents with cohabitation experience were asked to list all partners with whom they had cohabited. For each cohabiting partner the respondent named, Add Health asked, “Are you currently cohabiting with {initials}?” and “Is {initials} male or female?” If the respondent reported any current different-gender partners with whom they were cohabiting, they were classified as currently cohabiting.

As in the case of Add Health, there are two strategies to identify cohabitors in the NLSY-97: roster and survey questions. The roster method identifies cohabitation based on individuals who report “lover/romantic partner” on the household roster. The survey includes relationship questions beginning with the following prompt: “In this study we define a marriage-like relationship as a sexual relationship in which partners establish one household and live together.” Respondents reply whether they were in a marriage-like relationship. Using responses to these questions and information regarding past marital history, the NLSY-97 created a variable indicating marital and cohabitation status as of the survey date for each round. Possible responses include “never married, cohabiting,” “never married, not cohabiting,” “married, spouse present,” “married, spouse absent,” “separated, cohabiting,” “separated, not cohabiting,” “divorced, cohabiting,” “divorced, not cohabiting,” “widowed, cohabiting,” “widowed, not cohabiting.” Respondents are considered to be currently cohabitating if they report any cohabitation in the responses. The gender of the cohabiting partner(s) over the past year is established to ensure that it was a different-gender partner.

The NSFG current cohabitation status is based on a single-item question. Respondents are shown a card to identify their marital or cohabiting status, and the category not married but living together with a partner of the opposite sex is coded as currently cohabiting. The other relationship options include “married,” “widowed,” “divorced,” “separated, because you and your spouse are not getting along,” and “never been married.” These are not necessarily mutually exclusive categories: respondents could be cohabiting but separated (Gates 2011).

In the CPS, the household head reports on their relationship to all other household members and includes “unmarried partner” as the 13th of 15 relationship options. Individuals are included on the roster if it is their usual address. In addition, starting in 2007, unmarried household heads living with unrelated adults were asked whether they have a “boyfriend, girlfriend, or partner” in the household. If the answer is affirmative, the respondent identifies their cohabiting partner on the household roster. The same question is asked regarding all additional unmarried adults in the household—with the exception of members already identified as the cohabiting partner of the respondent—allowing cohabiting relationships that do not involve the household head to be identified. This strategy determines cohabitation among all household members and is called partner pointers. This moves beyond the basic roster strategy in the American Community Survey (ACS) or census of only identifying cohabitation of the head of household.

Cohabiting Experience

The cohabiting experience indicator captures whether the respondent had ever cohabited with a different-gender partner. The CPS can be used only to establish current cohabitation status and not cohabitation experience. In the Add Health, we draw upon the relationships section (discussed earlier) to determine cohabiting experience. If respondents named at least one different-gender cohabiting partner, they were defined as having ever cohabited.

In the NLSY-97, having ever cohabited is based on monthly reports of living arrangements. Respondents are asked at each interview, “Since the date of our last interview, have you been married to someone, or lived with a partner of the opposite sex in a marriage-like relationship where you established one household and lived together?” If such a living arrangement existed in the past year, respondents are asked to specify the month(s) during which such a living arrangement began and ended. The gender of the cohabiting partner was determined to ensure it was a different-gender cohabiting partner. Respondents who affirmatively reply to a cohabiting relationship with someone of a different gender in any month of the study are coded as ever cohabited.

The NSFG asked respondents about experiences living together with spouses and other individuals. Several questions that are used to establish cohabitation experience involve the wording, “Do not count ‘dating’ or ‘sleeping over’ as living together. By living together, I mean having a sexual relationship while sharing the same usual residence.” Respondents who are currently cohabiting or reported cohabiting with a spouse or other individual in the past are coded as ever cohabited.

Analyses

Our primary analyses were designed to identify the percentage of young women and men who were currently cohabiting or ever cohabited. The weighted percentages currently or ever cohabiting are estimated along with 95 % confidence intervals. The NSLY-97 is the reference survey in the tables, but more detail may be obtained from the upper and lower bounds of the confidence intervals presented in the online appendix, section B. We provide those estimates according to gender, education, and race/ethnicity. The sociodemographic indicators are measured in the same manner across surveys. Educational attainment was coded into four mutually exclusive and exhaustive categories: less than high school, high school graduate, some college, and a bachelor’s degree or higher. Race/ethnicity was coded as four mutually exclusive and exhaustive categories that are available in all the data sets: white non-Hispanic, black non-Hispanic, Hispanic, and other (including multiracial). Given the small sample sizes of Hispanics in the NSFG, we do not distinguish nativity status. Supplemental analyses indicate lower levels of cohabitation for foreign-born than native-born Hispanics. Gender was coded into male and female based on their response at the most recent interview. Age was measured at time of the survey and coded into three categories.

A standardization technique assesses whether differences in reports of cohabitation across surveys are due to sociodemographic compositional differences across surveys. This is important because surveys that disproportionately over- or underrepresented a subgroup (even with weighting), may be consequential for cohabitation estimates. Specifically, we ran logistic regression models of current cohabitation for the analytic sample of each survey that included indicator variables for the sets of variables displayed in Table 2 (i.e., education, race, gender, and age). Then we computed two sets of predicted values for the percentage cohabiting using coefficients from these models. One set utilized the sample means for each survey, and the other set utilized the sample means for the CPS. These sample means are displayed in Table 2. Comparisons of these two sets of predicted values indicate the extent to which the estimates are due to differences in sample composition.

Results

The weighted percentages and unweighted counts of respondents in our analytic sample are included in all tables. Table 2 demonstrates how the analytic samples differ according to education, race/ethnicity, gender, and age. With respect to educational attainment, the Add Health contains greater shares of respondents with some college education than the other data. Young adults in the Add Health and NLSY-97 have lower shares with only a high school diploma than the CPS.¹ The NSFG has a greater share of respondents with less than a high school diploma than the CPS. Approximately one-third of each data set is composed of college graduates. The race and ethnicity distribution differs across surveys, with a greater share of whites in the Add Health and NLSY-97, and a greater share of Hispanics in the NSFG and CPS than the other data sets. The gender and age distributions across data sets are similar.

Current Cohabitation Status

Figure 1 presents the percentage currently cohabiting along with 95 % confidence intervals. Overall, similar shares of respondents (i.e., about one-fifth of respondents) in all surveys except the CPS were classified as cohabiting at the time of interview. The levels are highest in the Add Health (22.6 % in the relationship survey questions) and lowest in the CPS (13.4 %). Figure 1 shows that estimates based on the CPS are statistically different than estimates based on the other surveys.² The NSFG estimates are significantly lower than the Add Health and NSLY-97 estimates. The Add Health estimates surpass the NSLY-97 estimates. The Add Health and NLSY-97 estimates based on the survey relationship questions do not significantly differ from their respective roster estimates. Given the comparatively large size of this age group, these relatively small percentage differences result in substantial differences in population counts. In terms of counts, more than 1 million more cohabitors are identified relying on the Add Health than the CPS.

We conducted several supplemental analyses to investigate the potential source of differences in the CPS and other surveys. To account for proxy reports of cohabitation in the CPS (household heads report on the status of other household members), we restricted the other data sets to household heads and found that this did not explain the difference (results not shown). We limited analysis to unmarried individuals to assess whether the CPS strategy of asking cohabitation status only of unmarried household members was driving differences. Again, this did not explain the gap in the CPS reports of cohabitation across surveys (results not shown). Additional sensitivity analyses demonstrated that differences were not due to the specific age range or period. The lower levels of cohabitation in CPS compared with the Add Health and NSFG persisted with the use of an older age range (29–31). (This contrast was not possible with the NSLY-97 because of the age restrictions of the initial interview.) Similarly, comparing the NSFG and CPS for a more recent period (2013–2015) also yielded significantly lower estimates in the CPS (results not shown). Further, we investigated whether number of months in the CPS sample, month of interview, or allocation flags explained differences in the CPS and the other surveys. These factors were not responsible for the lower levels of current cohabitation observed in the CPS. As another check, we computed the share of 26- to 28-year-olds in 2008 who cohabited based on a survey that relies on a similar roster and wording choice: the ACS. The estimate based on the ACS was 11.6 %, suggesting that the lower CPS estimates were largely driven by the roster and wording.

Table 3 presents the percentage of respondents who were currently cohabiting with a different-gender partner at the time of interview according to the sociodemographic indicators across the surveys. Significant differences between the estimates based on 95 % confidence intervals are presented in the online appendix, Table B1. For the sake of parsimony, we highlight significant differences between the NLSY-97 survey and other surveys. The CPS has significantly lower levels of cohabitation for each education category than the NSLY-97 survey as well as the Add Health and NLSY-97 roster (online appendix, section B). High school graduates and college graduates in the NSFG share similar levels of cohabitation as the CPS and lower levels than in the other surveys. An education gradient is apparent in all the surveys; the CPS education gradient distinguishes only the college-educated (see online appendix, section B). It appears that cohabitation among the least-educated may be the most seriously undercounted in the CPS.

Reports of current cohabitation status are significantly lower for each racial and ethnic group considered in the CPS than the NSLY-97 survey. White respondents in the NSFG report lower levels of cohabitation than in the Add Health or NLSY-97 survey but not different from the NSLY-97 roster or CPS. The levels of current cohabitation among blacks are significantly lower in the CPS than the NSLY-97 or Add Health. Blacks in the CPS and NSFG share similar levels of cohabitation. Black respondents in the Add Health report greater levels of cohabitation than in the NSFG and NLSY-97. Hispanic reports of cohabitation are similar across the Add Health, NSLY-97, and NSFG. Reports of cohabitation in the NLSY-97 survey and roster do not significantly differ for the racial/ethnic groups considered. However, within the Add Health, roster-based estimates of the levels of cohabitation among blacks are lower than survey-based estimates. These differences between the Add Health roster and survey suggest that it is how information on cohabitation is obtained (e.g., detailed question sequences and ACASI) rather than the sample that generates higher reports of cohabitation among blacks.

Men’s estimates of current cohabitation in the CPS are lower than in the other surveys. Women in the CPS, NSFG, and NLSY-97 roster report similar levels of cohabitation, and these levels are significantly lower than the NLSY-97 survey or Add Health survey or roster. Contrasts within surveys indicate that reports of cohabitation do not differ by gender (online appendix, section B). Each age group reports lower levels of cohabitation in the CPS than other surveys. The youngest respondents (26 years old) in the Add Health survey and roster report higher levels of cohabitation than in the other surveys. Among 28-year-olds, the NSFG has significantly lower levels than the NLSY-97 survey and roster but has levels similar to those of the Add Health and CPS.

Ever Cohabit

Figure 2 presents the percentages and the 95 % confidence intervals for cohabitation experience (shown in Table B2, online appendix). Levels of having ever cohabited are highest in the Add Health (69.6 %) and significantly lower in the NSFG and the NLSY-97 (63.0 % and 60.8 %, respectively). The Add Health estimates result in 1.5 million more young adults who ever cohabited than in the NLSY-97 (results not shown).

Table 4 shows the percentage of young adults who ever cohabited and demonstrates considerable variation according to sociodemographic indicators. (Table B2, online appendix, presents the upper and lower bounds of the confidence intervals.) An education gradient is once again evident, with cohabitation experience decreasing as education rises. At each education level, greater shares of Add Health respondents report having ever cohabited than NLSY-97 respondents. Respondents with the highest education levels (some college or college graduates) in the Add Health more often report having ever cohabited than similarly educated respondents in the NSFG. With regard to race/ethnicity, levels of cohabitation for Hispanic respondents are similar across the surveys. White respondents in the Add Health report greater cohabitation experience than white respondents in the NSFG or NLSY-97. Black respondents in the NLSY-97 report lower levels of cohabitation experience than their counterparts in the NSFG and Add Health. As a result, the racial/ethnic patterns in cohabitation differ across surveys. There are more differences in reports of cohabitation among males than females. Two-thirds of males in Add Health have ever cohabited, compared with 61 % in the NSFG and 55 % in the NLSY-97. Cohabitation experience is more commonly reported in the Add Health among 26- and 27-year-olds than in the other surveys. The differences in cohabitation experience are not significant for the oldest age category (28-year-olds) across surveys.

Standardization

As documented earlier, overall estimates of the percentage currently and ever cohabiting differ across the analytic samples from these surveys. On the one hand, this could reflect the fact that the demographic composition of these samples differs slightly, even after weighting. On the other hand, it could reflect the fact that estimates differ across surveys for specific demographic groups (e.g., respondents who did not graduate from high school). Given the patterns in the prior tables, we suspect that differences in cohabitation across the analytic samples of these surveys are driven more by survey features than by sample composition. We computed standardized estimates of current cohabitation for the Add Health, NSFG, and NLSY97 that constrain the demographic composition of these surveys to be identical to that of the CPS (e.g., Casper and Cohen 2000) (Table 5). A comparison of standardized values for the different analytic sample reveals the extent to which differences in estimates of current cohabitation are due to measurement as opposed to sampling.

The first panel of Table 5 shows that the estimates of the percentage currently cohabiting in Add Health, NSFG, and NLSY97 barely change when their distributions are standardized to match that of the CPS. Standardization increases the percentage cohabiting slightly for Add Health and the NLSY97 while decreasing it marginally for the NSFG. For example, the percentage cohabiting in Add Health increases from 20.15 % to 20.64 % with the switch from Add Health sample means to CPS sample means. Additional analyses (not shown) suggest that differences between the CPS and the other three surveys in the predicted percentage currently cohabiting are driven more by the intercept than the coefficients. Predicted probabilities for logit models estimated using sample means will not necessarily match observed probabilities (e.g., Cancian et al 2014), but it is reassuring that the predicted values for percentage currently cohabiting based on own sample means for Add Health, NSFG, and NLSY97 (rather than the CPS means) are extremely close to the values displayed in Table 2. We obtain similar patterns as presented in the lower panel of Table 5 when analyzing the share of young men and women who have ever cohabited using the CPS as the baseline estimates of composition of the age group (i.e., minimal shifts in the survey estimates). Importantly, the results from Table 5 suggest that differences across these surveys are largely an artifact of measurement rather than sampling.

Discussion

A critical task in family demography is to assess measurement of key family events, including cohabitation (e.g., Brown and Manning 2011). These data provide an opportunity to reassess the quality of data on cohabitation. Our study yielded four key conclusions. First, overall estimates of current cohabitation status and any cohabitation experience are different across the four surveys that we examine. We found that these differences appear to be a result of the measurement strategy and not the composition of the surveys. The CPS produced comparably modest estimates of current cohabitation status. Consistent with our expectations and findings from older studies (Bumpass and Lu 2000; Casper and Cohen 2000), reports from respondents of NLSY-97, Add Health, and NSFG produced higher population estimates of cohabitation than household head reports of cohabitation for all household members (i.e., proxy) in the CPS. However, this was unlikely the only cause of the differences because a gap persisted when we restricted the other surveys to heads of household. Further, these gaps persisted in the CPS and other surveys when we used a different age range or relied on a more recent time period. We believe the question wording (“unmarried partner”) and ordering (13th of 15 relationships) are the most likely explanations for the lower CPS estimates. The CPS is typically used to report levels and trends in cohabitation family living arrangements and may be underreporting current cohabitation status, especially for those with the lowest education levels and Hispanics.

Each survey with direct questions (Add Health, NLSY-97, and NSFG) differs somewhat in levels of cohabitation. Relying on the NSFG results in significantly lower levels of current cohabitation than the NSLY-97 or Add Health and significantly lower levels ever cohabiting than the Add Health. Respondents in the NSLY-97 and Add Health share similar levels of current cohabitation, but the Add Health respondents report significantly higher levels of ever cohabiting. These differences in cohabitation experience are likely most consequential for research predicting factors that are associated with entry and exits from cohabitation as well as implications of cohabitation for child and adult well-being.

More specifically, the question wording in the NSLY references “marriage-like” relationships, which may mean that only the most stable relationships will be defined by respondents as cohabiting. The lower levels of ever cohabiting in the NLSY-97 than the Add Health is consistent with prior work that shows questions about “marriage-like” questions yielded lower estimates of cohabitation than “ever cohabited” questions (Pollard and Harris 2007). Albeit problematic to change how relationships are measured in a longitudinal data, this approach should be reconsidered as this question wording persists in the current data collection.

The NSFG question sequence used to determine current cohabitation status is anchored around questions about marital status. This strategy means that NSFG respondents could be framing cohabitation as a more formal relationship status and must select whether they are currently separated or currently cohabiting; however, these categories are not technically mutually exclusive (Gates 2011). In contrast, the cohabitation histories in the NSFG are separate from the marital history questions and specify what does not constitute cohabitation, “dating,” or “sleeping over” and are quite specific about the nature of the relationship “sexual relationship while sharing the same usual residence.” Even though they adopt unique approaches, the NSFG and NLSY-97 obtain similar estimates in young adult cohabitation experience.

The Add Health provides the most detailed question sequences on cohabitation status, and along with its use of the ACASI, may explain the relatively high levels of cohabitation. Yet, higher levels do not mean they are the best or most accurate indicators of cohabitation. The higher estimates in Add Health may be due to the strategy of referencing each “romantic or sexual partner” to anchor questions about key events in the relationship rather than relying on one item requiring respondents to choose one status among many (e.g., married). Further, the question items were quite specific about the time frame of cohabitation (one month or more) and the independence of the residence (“neither of you kept a separate residence while you were living together”). This strategy requires more complex interview instruments and may not be feasible in general surveys. Our view is a strategy based on a single question that is akin to the Add Health item—“How many romantic or sexual partners have you ever lived with for one month or more? By ‘lived with’ we mean that neither of you kept a separate residence while you were living together.”—may provide a balance of a clear definition of cohabitation and a parsimonious approach.

Second, rosters and survey estimates of current cohabitation status within the same survey do not significantly differ, and this finding holds for each socioeconomic group considered in our study. The levels of cohabitation relying on the roster and the relationship questions more closely mirror one another in the NLSY-97 (95 % concordance) than in the Add Health (85 % concordance). These findings have implications for ongoing and future survey collections and suggest that there is not a serious problem with relying on rosters for identifying cohabiting couples. Thus, the CPS roster itself does not seem to be the sole reason why there are lower reports of cohabitation; rather, it may be the ordering of relationships (13th of 15 relationships) and the language “unmarried partner” used in the roster. All the rosters reference “usual” residence to ensure that all relevant individuals are included on rosters.

Third, we found variation in reports of cohabitation according to race/ethnicity, education level, gender, and age. The sociodemographic variation depends somewhat on the survey under consideration and whether current or ever cohabitation experience is being measured. One exception is that Hispanics report similar levels of currently cohabiting and ever cohabited across the surveys (except lower levels in the CPS). These lower levels in the CPS persist even when the analyses are limited to the native-born. A related issue is whether the same sociodemographic gradients exist within each survey. An education gradient is evident across all surveys (in the CPS, only for the college-educated), with lower cohabitation levels among the most-educated. The racial patterns in current cohabitation differ: in the CPS, significantly fewer blacks cohabit than whites; in the NLSY-97 and NSFG, similar shares of blacks and whites currently cohabit; and in the Add Health, more blacks cohabit than whites. No gender differences exist in reporting cohabitation within surveys except higher shares of women in the NLSY-97 claim to have ever cohabited. Thus, assessments of cohabitation differences for sociodemographic subgroups are not always consistent across surveys and are likely dependent on measurement. We found that the patterns for blacks and whites vary the most across surveys and merit greater attention. It is possible that findings related to magnitudes of sociodemographic differences in precursors and implications of cohabitation may differ across surveys.

A long-standing concern about cohabitation is whether it is a full-time or part-time living arrangement. The surveys attempt to clarify “usual” residence, but there are certainly part-time cohabitations along with relationships that are more ambiguous and in flux. Part-time cohabitation may occur as couples slide into and out of cohabitation, indicating a gradual and blurry transition (Binstock and Thornton 2003; Manning and Smock 2005; Pollard and Harris 2007; Sassler 2004). Knab (2005) reported that part-time cohabitation is relatively common (one in seven) among unmarried mothers. Using an earlier wave of the Add Health (2001/2002), Pollard and Harris (2007) found that 12 % of cohabiting women and 17 % of men had an additional separate residence. Further, cohabitors experience churning with periods of breaking up and getting back together, experienced by one-half of young adult cohabiting couples (Halpern-Meekin et al. 2012) and roughly one-quarter of cohabiting parents (Nepomnyaschy and Teitler 2013). The evidence shows relationships in more flux were more common among those reported by blacks, younger respondents, men, and lower-education groups. Thus, our work is consistent with prior research because it appears that cohabitation estimates among some population subgroups are more highly dependent on the measurement strategy (Knab 2005).

Fourth, compositional differences in the varying samples do not appear to be driving the differing levels of cohabitation across surveys. The standardized estimates suggest that the differences in the levels of currently or ever cohabiting among young adults are not due to the composition of the sample. It is notable that despite weighting, the composition of respondents differs across surveys. Thus, it appears the measurement of cohabitation rather than sample composition explains differences in the reporting of cohabitation.

Although our study provides an in-depth investigation of four major surveys, many other data collection efforts merit scrutiny. Attention to more recent estimates of cohabitation is warranted and may become more complex as men and women experience greater numbers of cohabiting unions (Eickmeyer and Manning 2018; Vespa 2014). Further investigation into cohabitation measurement in other federally sponsored surveys, such as the Panel Study of Income Dynamics, National Health Interview Survey, and SIPP, is warranted. Because cohabitation is certainly an increasingly common context for having and raising children, attention to measurement of cohabitation in surveys targeted at children is important, such as the Fragile Families, National Survey of Children’s Health, and Early Childhood Longitudinal Studies (ECLS-K, ECLS-B). A next step is to focus on how measurement issues influence research on the implications for children. Given growing levels of cohabitation among older adults (Brown et al. 2012), further attention to cohabitation measurement in surveys targeting their health and well-being is also important.

Our analysis is limited to one age group and one point in time, and we expect similar differences to exist across the life course. We believe that it is important to pursue additional research on comparisons of surveys over time. For example, Kennedy and Fitch (2012) provided an important contrast of cohabitation with and without partner pointers. Our work is limited to different-gender couples because the NSFG items are restricted to different-gender relationships and therefore exclude same-gender cohabiting couples. Comparisons of the measurement of cohabitation among same-gender couples across surveys are an important next step. These will be possible with new relationship status options and ordering in the census, ACS, and CPS with “opposite-sex spouse,” “opposite-sex unmarried partner,” “same-sex spouse,” and “same-sex unmarried partner” as the first four relationship options. Finally, testing how responses to questions on cohabitation differ according to education level, racial/ethnic groups, and gender is warranted.

Researchers must develop valid survey instruments that can produce accurate estimates that capture diversity in family dynamics. Unfortunately, there is no gold standard for measurement of cohabitation, such as a federal registry or administrative data. We presented variation in the measurement of cohabitation across nationally representative surveys and suggest a more uniform strategy to measure cohabitation and the inclusion of all unions, not just different-gender unions. Even though cohabitation is widespread, it remains an incomplete institution, with resulting measurement challenges, including blurred lines about the starting and ending of cohabiting unions (e.g., Avellar and Smock 2005; Binstock and Thornton 2003; Manning and Smock 2005; Pollard and Harris 2007). The inclusion of cohabiting relationships that are in flux or churning are important and may be better measured with specific direct questions than rosters. Our findings have implications not only for the design of national surveys but also for the interpretation of results based on the surveys that we compared. Attention to differences in sampling and questionnaire strategies are important factors in producing accurate estimates of cohabitation as well as correlates predicting the formation and stability of cohabitation along with research on the implications of cohabitation for children and adults. Researchers need to be more aware of the limitations and benefits of each strategy to collect data on cohabitation and to specify how cohabitation is measured. Casper and Cohen’s (2000:245) conclusion is still on point: “[W]e must consider more carefully how cohabitation ought to be conceptualized and whether it should be conceptualized different across surveys, depending on the purpose of the study.” We urge family scholars and policy makers to consider these implications when interpreting results about cohabitation across data collections.

Acknowledgments

This research was supported in part by the Center for Family and Demographic Research, Bowling Green State University, which has core funding from the Eunice Kennedy Shriver National Institute of Child Health and Human Development (P2CHD050959). This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by Grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due to Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health data files is available on the Add Health website (http://www.cpc.unc.edu/addhealth). No direct support was received from Grant P01-HD31921 for this analysis. An earlier version of this article was presented at the 2017 annual meeting of the Population Association of America. We appreciate helpful comments provided by Karen Benjamin Guzzo and Krista K. Payne.

Notes

References

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.